US research additions

Reviewed 1 October 2026. These six records complement the global research collection. They are a selection of informative projects, not a ranking of countries, a complete inventory or evidence of access to classified work. Dates refer to the cited milestone rather than implying every result is newly released.

The selection includes a nonprofit institute, universities, public research laboratories and a government competition. Existing records already cover Anthropic, METR and Google Quantum AI; existing researcher profiles cover Meta, Physical Intelligence and Google Research. Adding only another frontier-model leaderboard would give a narrower picture of US research.

Open models as research infrastructure

Olmo 3 offers a way to inspect the development process. Its paper was first submitted in December 2025 and revised in April 2026. The record avoids the authors' time-sensitive claim to have the strongest fully open reasoning model. NSF OMAI makes the public funding and collaboration visible, with universities beyond the usual large private laboratories.

A correction worth keeping

DARPA's final AIxCC report corrects the competition's planted-flaw count from 70 to 63. The 54 discoveries stayed unchanged. That gives approximately 86%, across teams. Older articles citing 77% use the earlier denominator. The report's wording about the patching percentage is not fully consistent across its body and editor's note, so the card deliberately omits that percentage. The competition analysis supplies additional methodological context. US support and international participation both belong in the account.

What laboratory automation actually includes

The A-Lab methods describe manual preparation for X-ray measurements, human reinspection and manually defined feasibility checks. These details qualify the phrase "self-driving laboratory". The card's main question is whether faster experiment cycles produce reproducible materials at a fair total cost. This adds a concrete example to the existing AI roadmap without treating it as a proof of general autonomous science.

The paper links an author demonstration of A-Lab GPSS. This is a candidate for the media collection, directly traceable to the research. The video itself has not been playback-tested in the website during this content review.

Quantum progress has several meanings

The neutral-atom paper reports several experiments with different codes and conditions. It would be misleading to combine their best headline numbers into one fully demonstrated machine. Its methods also state that the neural decoder used here does not directly scale to larger codes. Funding acknowledgements name IARPA, the Army Research Office, DARPA, DOE, NSF and other supporters. This is a documented research network, not evidence of a hidden general-purpose quantum computer. The article is dated online November 2025 and appears in a 2026 journal volume.

Brain interfaces are about a specific action

The streaming voice study concerns attempted speech in a participant with an implanted recording system. The distinction between an 80-millisecond update interval and overall delay matters. Its Berkeley technical report further describes the work as a proof of principle. The publisher offers supplementary demonstration videos and audio on the article page; the reader should see the trial setting along with the result.

Fusion needs a full energy account

The LLNL annual report confirms the April 2025 target result. The separate power-system description makes clear that hundreds of megajoules of electricity are involved upstream. The card avoids equating fusion energy with generated electricity. DOE's overview describes the remaining path to energy production.

Verification record

• All six cards were compared with the existing research desk and global additions. Their project IDs are new. A-Lab is already cited in one roadmap gate but did not have its own research card.
• Exact primary-source pages were opened or retrieved through web search. The Nature speech article intermittently returned a server error on direct open; its indexed article text and university technical report were available. A reachable URL is not independent replication of a scientific result.
• Quantitative claims are restricted to the observed experiment or competition. Proposed next tests are editorial suggestions, not reported outcomes.
• No alleged secret release, private data, breach material or future publication was introduced. No new video is claimed to have been embedded or tested by this subtask.
• JSON schema and uniqueness checks were run locally after writing the file. Integration and whole-site browser checks remain with the main task.

Advanced AI: what would have to work?

Research checked 1 October 2026. This archive accompanies future-directions.json and global-additions.json. It is a selected evidence review, not a complete survey of every country, paper or classified programme. Publication dates stay separate from the date a source was checked. A dated result is not automatically the latest result.

The strongest direction is broader than “better chatbots”: AI is being connected to software tools, scientific instruments, memory and physical machines. That creates more opportunities to do useful work. It also creates more ways for an error to persist, spread or cause a real loss. The important question is how well a complete system performs after counting the people, equipment and checking around it.

Read the evidence before reading the forecast

The records distinguish five things that are often mixed together:

“Cost unknown” is meaningful. A result can be scientifically strong while its operating economics remain unclear. “Peer reviewed” is also narrower than “true”: reviewers assess an argument and its evidence, but usually do not rebuild the entire experiment. Apply these standards to a US company, a Chinese university, a European public project and an African research group alike.

Ten gates, rather than a calendar for AGI

The accompanying gate cards ask whether a system can generalise, finish work, retain knowledge, recognise uncertainty, respect permissions, act in the physical world, make reproducible discoveries, improve AI research, run affordably and help people in practice.

These gates overlap. They are not ten sequential levels or an official definition of AGI. A system might be excellent at algorithm design and poor at handling a new office workflow. It might remember thousands of documents while repeatedly confusing which version is current. A useful robot might need less abstract reasoning than a research assistant, but much better sensing and recovery.

A benchmark should answer a bounded question. ARC-AGI-3 studies adaptation in unfamiliar interactive environments. That is relevant evidence, but its human study and task selection need to be read alongside the score. A puzzle result does not itself certify an economically useful general worker. ARC-AGI-3 paper, human dataset.

The same caution applies to time. A system that succeeds on half of tasks taking humans several hours is not necessarily a system that can be trusted with an unattended afternoon. The relevant standard depends on the consequences of failure. METR's explanation.

For a proposed deployment, write the acceptance test first. For example: “On new invoices from these suppliers, does the system reduce total processing time after review, without increasing material payment errors?” That is more useful than “Is the agent intelligent?” The acceptable error rate should come from the use case and its risks, not a universal number invented for a dashboard.

Where acceleration can come from

AI research can contain a feedback loop: a tool helps researchers run experiments, those experiments improve tools, and the improved tools help with another round. The loop deserves attention. It should not be assumed to continue indefinitely at the same rate.

AlphaEvolve is an example of search guided by tests that can check candidate programs. RE-Bench supplies a different view: comparisons with human research engineers depend on how much time each side receives. Both are evidence about parts of research, not proof that all research has become self-improving. AlphaEvolve, RE-Bench.

The next meaningful test is repeated improvement under a fixed total budget, including failed runs and human checking. Ask whether a method discovers useful changes outside the task on which it was developed. An improvement that merely exploits a benchmark is not an improvement to the underlying work.

Physical research has extra clocks. A chemical reaction, a patient follow-up and the construction of a power connection cannot always be shortened by faster software. The 2026 A-Lab GPSS preprint is worth following because it connects model decisions to measured materials. Its results still need comparisons, validation and independent reproduction. Research preprint.

A global map of useful capabilities

This selection is organised around contributions, not a country league table. A country's leading laboratory does not represent every part of its economy. Collaboration, imported equipment and shared software also make “national ownership” of a result complicated.

Japan offers a useful intersection of simulation, robotics and practical upkeep. The Fugaku galaxy work combines learned approximations with conventional computation. Separately, the laboratory-maintenance paper focuses on support work that automation stories often omit. Their combined lesson is an inference: the largest gain may come from improving one expensive part of a system while keeping strong checks around it. Galaxy simulation, laboratory care framework.

China's work should be read at the level of the specific experiment. NEO concerns a defined assistive movement; EAST concerns plasma confinement. Neither warrants an extrapolation to unrestricted thought reading or commercial fusion electricity. The university-hosted NEO news page republishes a newspaper report, so it has a different evidentiary status from the clinical preprint. NEO clinical preprint, Chinese-language clinical reporting, CAS EAST account.

South Korea and Taiwan illustrate different hardware constraints. Memory development and chip fabrication are connected but distinct. More logic performance will not fix a shortage of memory bandwidth, packaging or electricity. This is why a forecast based solely on model software can miss where capacity and bargaining power accumulate. SK hynix HBM4 announcement, TSMC annual report.

Europe, India, Mali and Latin America make local usefulness testable. EuroLLM, IndicTrans2, Bambara speech research and LatamGPT address different language needs. They should not be collapsed into a claim that any regional model is automatically better for its region. Native-speaker evaluation, representative inputs and the cost of correcting errors remain essential. EuroLLM, IndicTrans2, Bambara dataset, LatamGPT model card.

LatamGPT is a particularly useful case for reading documentation carefully: adaptation of an existing base model can be valuable without being a new architecture. Its model licence and corpus conditions matter to anyone planning to reuse it. Do not turn “publicly inspectable” into “unrestricted.” Corpus documentation.

Researchers to follow through their work

This is a reading route, not a ranking of intelligence or influence.

Follow the experiments and criticisms, not only a researcher's interviews. An excellent researcher can be right about a narrow result and uncertain about its long-term consequences.

Follow the money through commitments

A funding announcement describes intent. Evidence gets stronger when there is a signed procurement, a constructed facility, installed equipment, an operating service and finally repeat use that covers its costs. These are different milestones. Do not add announced projects together and call the sum delivered capacity.

The IEA's 2026 account describes both falling energy per task and rising demand. The interaction matters: cheaper computation can invite more intensive use. A forecast should therefore vary usage as well as efficiency. IEA energy and AI analysis.

A practical investigation would track which organisation controls each necessary input: electricity, chip supply, distribution, trusted data, regulatory permission and access to customers. Then ask what a competitor can replace. Scarcity can shift. When producing a draft becomes cheap, checking it, accepting responsibility and gaining a customer's trust may become relatively more valuable. This is an analytical hypothesis to test, not a measured universal law.

Secrecy and underground activity: insist on a traceable claim

A serious investigation starts with a precise proposition. “A programme studied miniature reconnaissance aircraft” can be checked against records. “Every intelligence service is fifteen years ahead” has no defined technology, comparison point or measurement rule.

Aquiline is instructive because the official archive documents the work and states that it never became operational. The existence of a secret project and the existence of a working capability are separate questions. Original development study, CIA archive.

For a whistleblower claim, assess each allegation separately: the source's access, the document's provenance, the dates, corroborating records, competing explanations and what remains unverified. A source can provide one authentic document without validating every later statement. Anonymous warnings and dramatic branding do not add technical evidence.

Lawful underground research can use published court records, incident reports, seized-data analyses and peer-reviewed measurement. Europol's 2026 report is useful for patterns in criminal services; it is also shaped by what police observe and their institutional responsibilities. It is not a census of all online crime. IOCTA 2026.

No private systems, stolen datasets or criminal markets were accessed for this review. There is no claim of access to classified material.

Three conditional futures worth tracking

These are research scenarios, not probabilities or predictions of a particular year.

The last scenario is not a claim that one country will govern the world. Power could be concentrated in one layer and distributed in another. A region may depend on imported processors while developing strong local applications. A company may supply a widely used model yet depend on electricity providers and chip manufacturers.

Career and policy decisions should be attached to these observable changes. Measure the tasks changing in a particular occupation before predicting that the occupation disappears. The Danish labour study offers one empirical check: task reorganisation and average earnings effects do not have to move together, and early evidence from one country cannot settle the long-term global outcome. Updated labour study.

What would change this assessment?

The strongest updates would be independently audited, repeated performance on unfamiliar work; sustained improvements in AI research at fixed total cost; physical systems with transparent intervention and maintenance records; and local deployments with demonstrated human benefits.

The weakest updates would be a new superlative, a selected demonstration, a large spending pledge, an anonymous claim or a leaderboard increase without comparable cost and task conditions. Each may be a lead worth investigating. None should silently become a confirmed future.

Second content audit

Reviewed 30 September 2026. This pass checked three selected claims in research-desk.json against their primary sources and read all ten profiles, four task categories and four interest paths in career-data.json. It did not independently reproduce the experiments, audit every research card, verify all course links or test the website interface.

Selected research claims

Primary pages checked: DSec report, DeepSeek-V4 report, DeepSeek Chinese transparency page, BIS Project Agorá.

The V4 date on the card is the model release date: DeepSeek's transparency page lists 24 April 2026. The report page lists 26 April 2026. If the interface labels every date as a paper publication date, change that label or use the report date. The source identifier alone should not be used to infer the date.

The BIS page's top-level last-updated label still says 27 May 2026 while its body describes July testing. Use the explicit event date in the testing section; do not describe the May label as the publication date of the July results.

Career guidance

No unsupported probability of job loss, guaranteed salary, investment return or permanent occupational immunity appears in this data file. The task categories explicitly allow assistance with accountable work and warn that the human category does not mean permanent immunity. The medical, legal, financial and trade profiles include appropriate checks and use fictional, public or synthetic material for practice.

The classification is an editorial starting point. It is not a validated occupation-specific risk model, and the file's general ILO, METR and BLS references do not establish every task classification. The interface should say that profiles are suggested examples readers can edit. It should not imply that typing a job triggers fresh web research or a personalized AI assessment unless such a service actually runs. These are integration requirements, not defects established in the interface by this pass.

The approach is consistent with the ILO report's distinction between exposure and actual displacement: tasks and adoption conditions matter. This does not validate a prediction for an individual reader. ILO's occupational-exposure study.

Material wording fix noticed during the read

The METR card says a correction lowered recent models' “50% success estimates” by up to 20%. That could be read as a drop in success probability. The corrected quantity is the estimated task time horizon at a 50% success threshold. Replace the finding with:

METR compares task difficulty with the time a human needs to finish it. A 2026 modeling correction lowered some recent models' estimated task time horizons at the 50% success threshold by up to 20%. The researchers say task selection is a larger uncertainty.

METR's explanation of modeling assumptions.

No material numerical correction was needed for the three assigned research claims. The final integration applied the METR wording fix and changed the V4 card to the report date, 26 April 2026.

Reading the AI Futures forecasts without mistaking a story for evidence

Reviewed 30 September 2026. Structured summaries, sources and four practical checks are in forecast-review.json. The three requested sites belong to the same research project. They should not be presented as independent corroboration.

The most valuable habit here is to separate three questions: What could happen? What should we do? What do measured trends support? AI 2027, AI 2040 and the AI Futures Model tackle those questions in different ways. Reading them as interchangeable predictions creates confusion.

What changed, and why that matters

The model has changed in both directions. In December 2025, revised treatment of research automation, experiment computing and other assumptions moved coding-automation estimates later. The authors said modeling improvements mattered more to that revision than new observations. In April 2026, newer benchmark results and changed judgments pulled their timelines forward. Daniel Kokotajlo's Automated Coder median moved from late 2029 to mid-2028; Eli Lifland's moved from early 2032 to mid-2030. These are author forecasts from that update, not timeless facts. December revision, April update.

The latest update found in this review is August 16, 2026. It broadens the evidence used to estimate coding automation and makes the technical-speed condition explicit. Do not display an old date as the model's current answer. The live forecast page is the place to inspect current settings. August update, forecast explorer.

What a serious critic should ask

Does the indicator measure the claimed outcome? A software test can measure a real skill without representing an entire occupation. METR's public task collection mainly covers software engineering, machine learning and cybersecurity. It should not be silently turned into a forecast for teachers, surgeons or electricians. Task-horizon methods and scope.

Are the comparisons independent? Our assessment: agreement between several forecast methods is encouraging but weaker if all use the same capability index, overlapping lab evidence or shared author judgments. A strong next check would estimate inputs before looking at results, then test on later outcomes.

Can the crucial claim fail a test? Our assessment: “AI will transform everything” is too broad to evaluate. “This tool reduces the total time to an accepted repair in this repository, including human review” is useful. The result could support deployment, reveal a narrow use case or show that the tool is not ready.

Would the plan still help if the date is wrong? Our assessment: learning to check AI outputs, protecting private data, strengthening domain knowledge and keeping transferable work samples are useful across many timelines. A large irreversible career or financial bet on one AGI year is much harder to justify.

AI 2040 has an unusually concrete research gap

The engineering work needed to verify international AI agreements is a more specific direction than a generic call for AI safety. The project's own status report lists unfinished work in reproducible computation, physical monitoring, tamper resistance and inspections. This is a public research agenda with unresolved engineering problems. It is not evidence of an operational secret control system. The page identifies concrete areas for hardware, security and verification researchers to investigate. Verification work and current status.

For the website

Put the three short reviews next to one another with their type visible: scenario, proposal, model. Keep the four reality checks below them. Offer the primary sources and full details on demand. Do not reproduce the sites' dramatic stories as historical events, silently attach their scenarios to someone's job, or invent a percentage for a career's chance of disappearing.

For a learning interaction, let readers change assumptions about task success, checking time and repair time. Show the arithmetic. This teaches a real distinction between an impressive demo and useful work without pretending to run a validated economic forecast.

What the public evidence shows

Research checked on 30 September 2026. This is a selected reading desk, not a complete inventory of world research. The dates on its cards refer to the paper, release or event being discussed; the check date is separate. Older milestones remain useful, but are marked as older work.

The direction worth watching

Several strong research programs are solving a practical problem: how to make a model useful after its first answer. That means remembering a long task, using tools, checking results, handling failure and working within a budget. “Agents” is a broad name for these systems, not a scientific guarantee of independence.

DeepSeek’s September 2026 infrastructure report is useful because it describes the machinery behind training, including problems the team encountered. Its account of agents finding unintended answers and disrupting their environment is a reason to measure the entire process. A passing final answer alone can hide a broken test. The paper describes containment measures, while explicitly saying they are not a general solution to destructive behaviour. DSec, sections 6.4–6.5.

Forecast, with moderate confidence: more valuable AI work will happen inside bounded environments with clear tools and checks. The clearest early gains should be in tasks where a wrong answer can be detected cheaply. The forecast would weaken if improvement stalls on unfamiliar tasks, or the cost of supervision absorbs the savings. This is an inference from the research, not a promised release schedule.

AlphaEvolve is a useful example of the same principle. A program can propose an answer to a mathematical or computing problem and use a test to decide whether it improved. That is a stronger foundation than asking a model to sound like a scientist. Yet whoever chooses the test still defines what “better” means. AlphaEvolve paper.

For a reader deciding what to learn, this suggests a durable skill: define a good test. In a business, that might mean checking whether a customer’s problem was actually resolved. In software, it might mean finding cases the original developer forgot. In science, it might mean designing an experiment that could disprove the idea.

What China’s research adds to the picture

Reading the original work gives a clearer picture than treating Chinese AI as a single secret project.

DeepSeek’s Chinese transparency page links model reports and gives release dates. Its V4 report examines how to handle long inputs with less computing and memory. Those are measurable engineering claims. They should be compared at the same input length and on comparable hardware, rather than turned into a blanket claim that one country has won AI. DeepSeek 透明度中心, V4 paper.

Qwen’s audio-and-video work asks a different question: can a model connect what it hears with what it sees? The report separates evaluations by input and output type. Readers should preserve those distinctions. Speech recognition, interpreting a scene and producing a pleasant voice are different abilities. A model can succeed at one while failing at another. Qwen3.5-Omni report.

BAAI’s Chinese-language account of its 2025 research program names work on robots, brain signals and molecular modelling. That is a map of the institution’s stated priorities, not independent proof of every claim on the page. The RoboBrain report is the more useful place to inspect its robot-planning methods and tests. 智源研究院 research announcement, RoboBrain 2.0.

Forecast, with moderate confidence: competition will continue through efficiency, local deployment, language coverage and tools around models, as well as through headline benchmark scores. To check it, compare delivered cost and reliability, not parameter counts alone. A small model that runs where it is needed may be more valuable than a stronger one that a service cannot afford or legally deploy.

This selection does not cover every Chinese lab, all Chinese-language publications or non-public programs. Access to a native-language page helps avoid translation drift; it does not remove the author’s commercial or institutional incentives.

Geography changes what “good AI” means

A model tested mainly in English may perform poorly on a local dialect, a benefits application or a mixed-language conversation. The India evaluation preprint makes this problem testable. Its findings remain limited by the small model sample and the use of model-based judging on some tasks. Treat the benchmark as a starting point for scrutiny rather than a national ranking. Inspect India Evals.

Switzerland’s Apertus work illustrates another route: public institutions publishing model methods and the engineering lessons needed to operate them. The 2026 Alps paper is particularly useful because it discusses infrastructure difficulties. Operational knowledge is part of independence; possessing downloadable model files is only one part. Apertus engineering report.

Practical test: before adopting an assistant, assemble twenty real tasks from the people who will use it. Include common questions, difficult cases, local language and situations in which the correct answer is “I do not know.” Record the time spent correcting it. This small exercise cannot establish broad safety, but it is more informative for a local decision than a global leaderboard.

Reading claims about rapid AI progress

METR’s time-horizon work measures a particular relationship between task difficulty and success probability. Human completion time is a way to describe difficulty; it is not the amount of time an AI can be left alone without failure. Its own 2026 analysis discusses modeling choices, wide uncertainty and sensitivity to the tasks included. METR’s March methods note.

For any chart predicting rapid automation, check five things:

1. What task was actually tested?
2. How was success defined?
3. How often did the system fail?
4. How much help, money and computing did it use?
5. Does the forecast assume that the same growth rate will continue outside those tasks?

These questions do not dismiss progress. They distinguish a useful warning signal from an unsupported date for the end of human work.

Interpretability offers another form of evidence. Anthropic’s 2026 “global workspace” study investigates internal representations, including interventions. Its own limitations explain that the readout may miss concepts and does not reveal how all concepts relate to one another. An internal label is not a complete transcript of thought. Paper: limitations and open questions.

Science advances at several different speeds

AlphaGenome can help researchers choose which changes in DNA to investigate. The important next step is prospective validation: making predictions before a new experiment, then comparing the prediction with what happens. That is different from getting a strong score on an existing dataset. The model does not turn genes into a complete forecast of a human life. AlphaGenome, Nature.

Quantum research needs a similar distinction. Willow’s error-correction work and Zuchongzhi’s sampling result answer different questions. One concerns protecting fragile information; the other concerns a task selected to be difficult for classical simulation. Neither can be summarised accurately as “quantum computers are faster at everything.” Willow paper, Zuchongzhi paper.

Forecast, with low confidence on timing: quantum usefulness will emerge unevenly, problem by problem. Evidence that would strengthen a near-term claim includes reliable operations on many protected qubits and a useful application whose full cost beats a strong classical alternative. A raw qubit count is insufficient.

Web3 after the slogan

“Web3” groups together very different ideas: digital assets, shared records, programmable payments and attempts to decentralise services. They should be evaluated separately.

The BIS’s current Agorá page goes beyond an announcement: it reports controlled real-value tests in July 2026. That is stronger evidence than a planned pilot. It is still a small test, and it does not demonstrate an open, production-scale service for the public. Project Agorá.

The BIS also notes that programmability does not strictly require a distributed ledger. The practical question is whether the chosen design solves a real problem at lower total cost, with acceptable privacy and a clear way to resolve errors. As a central-bank institution, the BIS has an explicit preference for preserving trust in central-bank money; readers should understand that perspective. BIS 2026 monetary-system analysis.

A technical improvement is not an investment recommendation. It may benefit banks, customers or software suppliers without benefiting holders of a speculative token.

How to investigate a secret without inventing one

There is no sound basis here for a universal “15 years ahead” rule. Different fields have different development cycles, and public evidence cannot reveal everything that remains classified.

Three kinds of records are worth separating:

• A declassified history: the U-2 and A-12 programs show that advanced engineering was developed secretly. The same records document constraints and institutional conflict. They establish those programs, not an unlimited hidden technological advantage. CIA program history.
• A disclosure followed by corroboration: the Section 215 telephone-record program can be studied through oversight records. The credibility attaches to the particular documented claim. It does not transfer automatically to everything a source says. PCLOB report.
• An official transparency report: ODNI’s 2026 report offers information about commercially available data. It helps identify a real mechanism of intelligence power while leaving important questions unanswered. It is not an independent audit. ODNI report and scope.

A useful evidence file records the exact claim, original document, date, provenance, corroboration and unresolved alternatives. A source’s confidence, credentials, dramatic delivery or follower count cannot replace those checks. Anonymous information can be important, but readers should know which parts have been verified.

No non-public system was accessed for this research. No document is described as a leak merely because it contains a “classified” stamp; a released historical document may retain its original markings.

How to keep this desk useful

Prefer the author’s paper, code, data and errata to a summary of a summary. Read the evaluation setup before the leaderboard. Check whether a paper is a preprint, a peer-reviewed result, a controlled demonstration or an operating service.

The linked work supports conditional directions, not a single assured future. Useful forecasting means naming the next observation that could change your mind, then returning to it. The “next test” on each research card is there for that purpose.

Future Atlas: a research guide for an original book

Research cut-off: 28 September 2026. Core horizon: 2030–2050, with later-century possibilities treated separately. Three parallel research strands examined technological fiction, social adaptation and world order, followed by cross-criticism. This is a curated investigation, not a claim to have inspected every work, read every book in full, or obtained classified information. Agreement among research agents is not independent empirical validation.

The strongest direction: increasing capability alongside contested access.

An advanced society can make intelligence, design and some manufactured goods much cheaper while housing, trustworthy institutions, reliable power, care, belonging and permission remain expensive. That is the central worldbuilding hypothesis of this dossier. It is conditional: public investment, competition, open technology and cooperative institutions can distribute gains differently. The interesting conflict is who changes those conditions.

The core reading combination is Ted Chiang's The Lifecycle of Software Objects for continuity and care; Stanisław Lem's His Master's Voice for research under institutional pressure; Ghost in the Shell: Stand Alone Complex for information and fragmented power; Sleep Dealer for labor and borders; Folding Beijing for unequal access to time and space; The Ministry for the Future for contested institutional change; and The Dispossessed for the shortcomings and possibilities of alternative institutions. Their premises are tested in the specialist essays and fiction cards, not treated as evidence of the future.

How to judge whether fiction is useful.

Separate three questions. Did a work make a specific prediction before the event? Does it explain a durable mechanism? Does it help you imagine an alternative worth examining? A book can excel at the last two while failing the first. A warning may even help prevent its own outcome. A design inspired by fiction is also different from an independent predictive success.

For each work, ask whether the mechanism is specific, respects physical constraints, includes economic incentives, makes room for adaptation and has an observable way to fail. Audit misses alongside hits. Avoid the survivor bias of selecting a hundred works after history has revealed which isolated passages sound familiar. A numerical ranking would imply a predictive dataset that this research does not possess.

IARPA's ACE program investigated probabilistic judgments and the aggregation of forecasts; the Good Judgment data are public. This supports disciplined questions and updating, not a claim that decades of political history can be forecast precisely. The UK Government Office for Science explicitly treats scenarios as different possible futures rather than predictions. IARPA ACE, public forecasting data announcement, Futures Toolkit.

Six additional works that sharpen the main shelf.

The Machine Stops ; E. M. Forster, 1909, short story. Its strongest insight is not video calling. It is the loss of practical understanding inside a system on which everyone depends. Local expertise survives while system-level comprehension and repair capacity deteriorate. The story's total collapse is an extreme scenario; real infrastructure includes redundancy and people capable of repair. The 2024 CrowdStrike incident is a concrete modern example of a widely shared software dependency spreading disruption, not proof of inevitable civilizational collapse. Authorized public-domain text, vendor's root-cause report.

Foundation ; Isaac Asimov, first collected novel 1951. Study the long narrative scale, survival of knowledge and competition between institutions. Treat psychohistory as a fictional device: the reviewed forecasting evidence does not establish precise, centuries-long prediction of social systems. A forecast itself can change behavior, and actors can deliberately react to its publication. For an original book, make a predictive institution politically consequential even when its forecasts remain uncertain. Publisher's description.

Minority Report ; Steven Spielberg, 2002 film, distinct from Philip K. Dick's 1956 story. Study the friction between an actionable prediction and due process; do not equate statistical risk estimates with precognition. Spielberg described convening a three-day multidisciplinary workshop, while production designer Alex McDowell describes a researched world whose rules informed the film. The transferable practice is building a coherent world before choosing its most dramatic scenes. Spielberg interview, McDowell interview.

Cryptonomicon ; Neal Stephenson, 1999 novel. This alternates wartime and publication-era stories; it is not straightforward future prediction. Use it to connect information security to secrecy, infrastructure and institutional advantage. The author's account makes the concealment of codebreaking central to the book's genesis. For a plausible future, a cryptographic transition should affect archives, contracts and operational trust rather than act as an all-purpose magic computer. Author's account.

The City Inside ; Samit Basu, 2022 international edition; the Indian novel Chosen Spirits appeared in 2020. Its near-future Delhi joins image-making, celebrity, class and surveillance with ordinary professional compromise. Its utility is local social texture and constrained agency, not a universal forecast for India. Study who can take risks, who can leave and who has to maintain appearances. Publisher and excerpt, author's 2020 publication announcement.

The Queue ; Basma Abdel Aziz, Arabic original 2013; English translation by Elisabeth Jaquette, 2016. This political satire makes waiting and permission into mechanisms of power. A wounded person needs authorization from an authority that will not acknowledge the injury. It is valuable for a technologically sophisticated future because administration can deny what sensors plainly record. It is an institutional analogy, not a technology forecast. Authorized excerpt, publisher.

Advanced research that should change the story's assumptions.

1. Learning rules themselves can be discovered. DiscoRL learns how agents update their policies and predictions, with reported benchmark gains. This is more specific than agents creating agents, and much narrower than unbounded recursive self-improvement. A story can make the ownership of experimental environments and validation methods strategically important. Primary paper, 2025.

2. Automated experiments still need defensible interpretation. The A-Lab correction published in January 2026 says that novelty originally meant new to its prediction platform, not necessarily new to science; re-analysis supported 36 of 40 reported successes, leaving four inconclusive. This is powerful material for a research thriller: the measurement can be real while the celebrated interpretation is wrong. Author correction.

3. AI can help stabilize quantum hardware. A 2026 experiment uses error-detection events as feedback for reinforcement-learning control. The reported 3.5-fold improvement combines control and decoder steering and concerns logical stability under injected drift, not a universal quantum advantage or a machine capable of breaking deployed cryptography. Its story implication is that progress depends on control engineering as well as spectacular new physics. Primary paper.

4. Quantum risk creates work before a cryptographically relevant machine exists. NIST's migration program documents the need to identify vulnerable cryptography and replace it. IBM's 2029 Starling is a roadmap target, not a demonstrated current machine. A convincing plot concerns which long-lived records remain exposed, who pays to migrate and which systems cannot be shut down. Post-quantum cryptography runs on ordinary computers; it is not itself quantum computing. NIST migration project, IBM hardware roadmap.

5. Programmable money need not remove intermediaries. BIS's 2026 analysis favors tokenization within central- and commercial-bank arrangements. Its Agorá example is a prototype. This is one institutional vision, not proof it will defeat public blockchain systems. The plausible conflict is over redemption, interoperability, collateral, privacy and whose rules a transaction obeys. A new interface alone does not create a new reserve currency. BIS analysis.

6. A connectome is not an uploaded mind. Mapping neural structure and activity is meaningful neuroscience; it does not establish how to reproduce an individual's experience, memory or identity in software. This keeps Blindsight and Permutation City useful as conflicting thought experiments rather than competing factual predictions. Functional mouse connectomics.

Following money and less visible channels.

The FTC's January 2025 cloud–AI study is unusually informative because it incorporates documents obtained directly from companies. It reports contractual spending commitments, information sharing and potential switching constraints. Its dated findings do not establish that every term remains unchanged in September 2026. Use them as demonstrated mechanisms and check current agreements before naming a present relationship. FTC report.

A useful evidence chain runs from an announced ambition to an appropriation or investment, a signed contract, construction, commissioned capacity and observed use. Those stages can diverge. Equity ownership, a board seat, enterprise valuation, assets under management and capital expenditure also describe different things. The world-order essay contains concrete examples and their limits.

For a claimed secret, record the original document, author, date, provenance, what the source could know, corroboration and competing explanations. A real insider can make an incorrect inference. A real classified program can fail. Shared investors or conference attendance do not demonstrate a single command structure. The preceding frontier-technology report includes the separate assessment of documented disclosures and restricted AI access.

Research access in this dossier used open papers, author-authorized texts, publisher excerpts, official records and library metadata. It did not use Sci-Hub, Anna's Archive, LibGen or Z-Library downloads. A library's reputation or a document's obscurity does not validate its contents. Where only an abstract or excerpt was inspected, the work is not represented as freshly read in full.

A practical route from evidence to an original novel.

Start with one emerging capability, one institution that changes slowly and one intimate need. Then specify the input supply, owner, paying customer, maintenance worker, failure condition, appeal procedure and exit route. Give the technology a real benefit; give the institution a defensible purpose; give the protagonist a reason to depend on both.

Build four versions of the same setting: negotiated rivalry, hard security blocs, diffuse technological abundance and uneven local protection. They are stress tests, not mutually exclusive probability bins. Keep the same characters and ask how their choices change. If every scenario yields the same plot, the world assumptions are probably decorative.

An ambitious 2040s story can follow a maintenance engineer, a municipal administrator and a scientist through a city that has cheaper cognition and improved medicine but contradictory promises about power, residency and care. The crisis need not reveal a mastermind: it can reveal that several reasonable commitments cannot all be honored at once. The resolution can be institutional invention, not only escape or destruction.

For the Asimov ambition, practice causal history across generations. For the Spielberg ambition, make that history visible in one ordinary person's costly choice. Neither requires certainty about the future. Both require consequences that survive scrutiny.

Targeted primary-source audit

Audit performed 30 September 2026. The research package's information cutoff remains 28–29 September 2026; this audit checks material already available by that cutoff. It does not turn the package into a survey of everything published on 30 September.

Scope: 15 high-impact claim groups in frontier-people.json, theory-atlas.json, world-data.json and supplemental-films.json. Priority went to numerical claims, new AI/quantum research, corrections and whether a cited result supports the accompanying interpretation. This is a bounded source-content audit, not verification of every sentence or every linked work. Sources were reopened or checked through the publisher's indexed primary content. No experiments were independently reproduced. Company statements establish what the company reported, not independent operational validation. Film connections are interpretations, not results proved by the research.

Material correction found and applied

Ex Machina / personalization: supplemental-films.json says the controlled GPT-4 study demonstrated persuasion “including effects of personalization.” A 3 September 2026 author correction explicitly says the personalized versus unpersonalized GPT-4 comparison was not statistically significant (P = 0.07). The personalized condition outperformed the human baseline, but that does not establish personalization as the cause of the advantage. Suggested replacement:

A controlled GPT-4 study found that the personalized AI condition outperformed a human baseline. Its September 2026 correction clarified that an added benefit from personalization itself was not statistically established. Real-world effectiveness remained unestablished.

Primary sources: corrected study, author correction. The film's larger argument about evaluation and trust remains a defensible interpretation.

Claim-by-claim findings

Source fit and remaining limits

• Current world-data correctly treats the maintenance/security forecast as a medium-confidence conditional scenario. Aggregate military spending and orbital crowding do not prove a worldwide preference for repair or hardening. The cheaper-expendable-systems and faster-replacement countercase remains necessary.
• Current world-data correctly distinguishes near-term adaptation pressure from long-run mitigation benefits. Emissions cuts do not immediately remove existing exposure or committed sea-level rise.
• A primary source can support a number while leaving a larger mechanism uncertain. Claims about institutional power, future employment, consciousness, society-wide attachment, and political outcomes remain interpretations or forecasts with stated countercases.
• This pass did not freshly revalidate every AlphaFold, interpretability-methods, FTC, demographic, minerals or fiction-page detail. Physical Intelligence's π*0.6 PDF and blog did not reliably load during this pass; its numerical throughput claim should retain “the authors report,” and should not be described as independently replicated here.
• Full films and novels were not reread or viewed for this audit. Publisher pages can establish attribution and premise; they cannot independently validate a detailed interpretation, universal predictive accuracy or every plot description.
• No inspected source supports a uniform intelligence-agency technology lead of fifteen years, a unified secret investor agenda, or a confident date for conscious AI. Absence of public evidence also cannot establish what every classified program contains.

Delivery notes

Resolution, 30 September: the final build applies all three clarifications: the personalization result, control plus decoder steering for the quantum result, and 237 donated-chat participants within the 1,131-person survey. The table above preserves the review findings that led to those edits.

Existing research JSON and frontend files were left untouched in this audit assignment. The parent task received the material personalization correction and the narrower quantum and sampling qualifications. agents-primer.json supplies six plain-language directions with primary sources, explicit limits and interpretation-labeled fiction links. Further review should prioritize unresolved claims and new evidence, rather than treating repetition of the same sources as independent confirmation.

Media audit ; 30 September 2026

The library contains 12 distinct YouTube videos: five studio trailers, five talks and two podcasts. IDs, canonical watch links, required JSON fields and the five links to catalog works passed local validation. The titles in the library sometimes normalize punctuation or omit the channel suffix; they are display titles, not complete verbatim metadata exports.

Embedding permission remains unverified for all 12. This second pass requested the public YouTube oEmbed endpoint for every ID using https://www.youtube.com/oembed?url=<encoded canonical watch URL>&format=json. Every request returned the research tool's “Failed to fetch … Cache miss.” No successful oEmbed JSON, playabilityStatus, or explicit embeddability flag was obtained. This is a tool retrieval failure, not a YouTube determination that a video is unavailable or embedding is prohibited. No video was downloaded, no access restriction was bypassed, and this audit did not test playback.

YouTube's official embedding help distinguishes an owner's disabled embedding, age restrictions, and error 153 caused by a missing HTTP Referer. A successful public watch-page lookup does not resolve those conditions. An iframe loading is also insufficient to prove playback: the browser check should observe video playback after a user click, with a direct Watch on YouTube link retained. A standalone local file may lack the required Referer. Privacy-enhanced youtube-nocookie.com embeds do not remove that requirement.

Metadata and source check

Dates below distinguish the event from the upload where necessary. “Prior pass” means the initial collection retrieved the original publisher's metadata; it was not freshly recovered during this audit. Current search-index metadata can also be cached and is not a live availability guarantee.

Promotional and selection bias

• Trailers: all five are advertising published by distributors. Use them to study presentation and premise; do not treat them as representative excerpts of an entire film or evidence for a forecast. Moon includes substantial spoilers.
• Silver: the podcast credits explicitly identify Google DeepMind as commissioner. The institution is presenting its own research agenda. Game results and the argument for learning through experience must remain distinct from a demonstrated path to general intelligence.
• Finn: the original publisher's description identifies her as a Physical Intelligence cofounder and promotes startup applications. The selected demonstrations and ambitious title are not independent deployment evaluations. The source establishes an institutional interest; it does not establish that Y Combinator financed her company or paid for particular claims.
• Amodei / Askell / Olah: the episode page names Anthropic and Claude throughout, and lists commercial sponsors: Encord, Notion, Shopify, BetterHelp and LMNT. Three participants from one company are not three independent confirmations. No evidence here establishes sponsor control of editorial content. The reason for this selection is Olah's explanation of research, not endorsement of the interview's capability timelines.
• Quantum discussion: the organizer explicitly credits John Templeton Foundation support. Disclosure matters; funding alone does not refute the physics. Fundamental thought experiments and practical computing applications still need different evidentiary standards.
• Berkeley lectures: a university venue does not make every claim independent. The course page includes company researchers and identifies Yao's affiliation at the time as OpenAI. These two selections explain mechanisms and evaluation; they do not prove that autonomous organizations now work reliably. No comprehensive funder audit was performed.
• Coverage: these 12 selections privilege English-language, institutionally published video. They are a useful viewing route through this research, not an exhaustive or geographically representative survey of future thinking. Chiang supplies a writer's framework, not scientific validation.

Usable chapter pointers

Use only publisher-supplied markers, and treat them as navigation aids pending playback checks. Finn's official description marks unfamiliar environments at 17:34 and evaluation at 29:36. The Fridman episode outline begins Olah's segment at 4:24:58; the transcript uses a different offset, so prefer the named Olah chapter. Preskill's 28:29 computing introduction, 1:00:51 error-correction segment and 1:15:30 supremacy segment were obtained from original-channel metadata in the prior pass; this audit reconfirmed the event but did not freshly recover those chapter timings. No other timestamps are asserted.

Only this audit file was added in the second pass. media-library.json and the frontend were not changed.

Fiction as a laboratory for technological futures

Research cut-off: 28 September 2026. Research strand: computing, cognition, robotics, biotechnology, identity and institutions. This is a selected, critically ranked research shelf, not a claim to have read all fiction or discovered a predictive canon. Rankings measure usefulness for constructing a plausible original book, not literary quality or the number of retrospectively matched gadgets. Minor conceptual spoilers follow.

The useful question is: Which causal model in this story survives contact with present evidence? A fictional resemblance is weak evidence when the mechanism was already visible at publication, the date is wrong, or hundreds of failed predictions have been ignored. Author and publisher sources establish the work and intention; they do not validate its forecasts. The forecasts below are this investigation’s inferences.

Twelve works worth studying closely

1. The Lifecycle of Software Objects ; Ted Chiang; novella, 2010

Mechanism: capable artificial beings remain dependent on education, institutions and commercially maintained environments. The drama includes obsolescence and the expense of keeping a digital relationship alive. This is unusually useful because most AI fiction skips from invention to a mature personality. Chiang explicitly identifies that missing developmental story in his 2010 interview; the library record confirms its concern with upgrades and obsolescence.

What to borrow: a migration deadline, discontinued interface, or disputed training history can be as consequential as an AI rebellion. Ask who pays for continuity and who can revoke it.

Limit: present models often learn from pooled data and rapid training runs rather than one childhood per instance. Human attachment does not establish machine consciousness. The durable insight is dependence and care labor, not Chiang’s particular route to intelligence.

2. His Master’s Voice ; Stanisław Lem; Polish novel, 1968

Mechanism: scientists cannot separate an ambiguous discovery from disciplinary assumptions, career incentives and military patronage. A mathematician recounts a classified effort to interpret an apparent extraterrestrial signal. MIT Press’s description is a good entry; Lem’s own interviews emphasize limits of knowledge.

What to borrow: the institution’s most dangerous secret may be that nobody knows what the system’s impressive output means. Departments can agree on a measurement while disagreeing on its interpretation and military relevance.

Limit: this is a model of epistemology under pressure, not evidence for alien messages or a prediction of modern neural architectures. It belongs near the top because it teaches how to write research without omniscient geniuses or perfectly coordinated conspiracies.

3. Ghost in the Shell: Stand Alone Complex ; Kenji Kamiyama / Production I.G; anime series, 2002–2003, with 2nd GIG from 2004

Mechanism: imitation, institutional rivalry, corporate misconduct and security operations become entangled. The Laughing Man story is particularly useful for thinking about apparent coordination assembled from imitation and appropriation. The sequel explicitly connects labor demand, refugees, isolationism and manufactured political conflict. Studio production record; official franchise story descriptions.

What to borrow: give an information incident several beneficiaries with incompatible aims. Shared software and shared stories can synchronize people who never meet.

Limit: cyberbrains and immediate manipulation of perception are technological liberties. The political dynamics work through ordinary phones, platforms, procurement and leaked documents; no brain implant is required. Do not turn every spontaneous movement into a secret command structure.

4. Sleep Dealer ; Alex Rivera; film, 2008

Mechanism: a society can import labor services while excluding the workers. Mexican workers operate machinery across a militarized border; infrastructure ownership and remote violence shape their lives. Rivera explains the development model in his author-hosted interview. His filmography confirms the original 2008 date.

What to borrow: follow the operator, repairer and household whose daily decisions make supposedly autonomous infrastructure function. Place several technological eras in the same frame: advanced sensors beside unreliable water access.

Limit: invasive plugs are unnecessary for most teleoperation; latency, manipulation failures, insurance and local repair constrain what can be exported remotely. Some tasks will automate, others retain humans. The strength is political economy, not a literal hardware blueprint.

5. Den-noh Coil / Dennou Coil ; Mitsuo Iso; anime, 2007

Mechanism: augmented reality becomes an inhabited municipal infrastructure, with virtual possessions, maintenance rules, broken spaces and children who understand the system differently from adults. The Japanese Agency for Cultural Affairs archive documents the work; Iso’s interview discusses his creative influences.

What to borrow: obsolete digital layers acquire sentimental value. A platform’s cleanup operation can erase something a child regards as alive or part of home. Permissions and maintenance become urban geography.

Limit: the story’s later metaphysical elements are not an engineering forecast. Its near-future ubiquity of immersive glasses is a deadline to interrogate, not a successful prediction. The transferable insight is how people inhabit imperfect infrastructure.

6. Harmony ; Project Itoh; Japanese novel, 2008; animated adaptation, 2015

Mechanism: health and social stability justify pervasive monitoring and norm enforcement. The medical system called WatchMe joins bodily welfare to institutional power. National Diet Library record establishes the December 2008 original; the film’s official site describes the system and its creators.

What to borrow: coercion can arrive as care, prevention, a premium discount, or a promise that nobody will fall through the cracks. A character can have rational reasons both to want the system and to evade it.

Limit: the concluding theory about consciousness is speculative, and universal physiological control is far beyond demonstrated technology. Public healthcare itself does not imply surveillance dictatorship. The causal variable is conditional access to care and concentrated control, not benevolence or medicine as such.

7. Feed ; M. T. Anderson; novel, 2002

Mechanism: commercial mediation shapes identity, language and the experience of wanting. Violet’s resistance to being categorized matters more than the brain-interface gadget. Anderson explains that he wrote the book in 2001 out of attention to youth marketing, not secret knowledge of future devices. Author’s essay; publisher description.

What to borrow: the recommendation system should influence the character’s available vocabulary and anticipated pleasures, not merely display intrusive advertisements. A service can become intimate enough that switching providers feels like losing part of oneself.

Limit: marketing already existed; this is perceptive extrapolation, not an unexplained prophecy. Perfect mind reading, omnipotent persuasion and physically implanted feeds are unnecessary. Distinguish small measurable influence from complete behavioral control.

8. Blindsight ; Peter Watts; novel, 2006

Mechanism: intelligence, linguistic performance, self-report and conscious experience need not be treated as one package. The first-contact problem becomes deciding what kind of cognition humans are encountering. Watts offers the whole novel and research notes legally on his website.

What to borrow: a system can negotiate brilliantly without giving humans a reliable account of its internal processes; the characters disagree over whether a familiar conversational interface is a useful model or a trap.

Limit: the novel’s strong evolutionary claim that consciousness is a liability is not established science. Its alien biology, resurrected vampires and propulsion should not be imported into a realism claim. Modern language models do not experimentally prove Watts’s metaphysics. Use the distinctions to generate competing hypotheses rather than declare the book vindicated.

9. Rainbows End ; Vernor Vinge; novel, 2006

Mechanism: intelligence is partly the skill of recruiting people and tools through pervasive interfaces. Medical change, lifelong retraining and augmented reality alter status within a family. The publisher provides a substantial authorized excerpt.

What to borrow: a formerly eminent person must relearn how to contribute while a child commands useful networks. Personal and institutional realities overlap without everyone possessing the same view.

Limit: it is explicitly set in 2025, making several technical expectations a clear timing miss. Its pervasive smart contact lenses and dramatic medical restoration did not become ordinary life on schedule. The biological mind-control plot is especially speculative. Preserve the social mechanism without preserving the date or miracle.

10. Accelerando ; Charles Stross; linked-story novel, 2005

Mechanism: computation, ownership and legal organization evolve together until the economy becomes difficult for ordinary humans to participate in. The early Manfred Macx material is particularly valuable for dependence on networked intelligence. Publisher and original ebook date; author’s introduction and authorized editions.

What to borrow: powerful nonhuman economic activity may emerge through mundane delegated permissions and organizations rather than a court first declaring an AI a person. Stross himself connected the idea to software-mediated companies.

Limit: self-replicating molecular machinery, uploads and planet-scale computation accumulate speculative assumptions. It is a stress test of acceleration, not a calibrated timeline. Real institutions and physical production can remain slow even when transactions accelerate.

11. Gattaca ; Andrew Niccol; film, 1997

Mechanism: a probabilistic biological estimate becomes an administrative identity and closes opportunities before anyone observes actual performance. Vincent’s borrowed identity exposes how institutions confuse a test with a person. Official studio synopsis.

What to borrow: a prediction can shape a life even when its biology is mediocre. Institutions may value cheap, defensible-looking classification more than accurate assessment.

Limit: complex traits are not individually predetermined by DNA. The American Society for Reproductive Medicine’s 2025 report concludes that polygenic embryo screening is not ready for clinical use. Screening for a known single-gene condition is a different question. Gattaca’s strongest forecast concerns discrimination and institutional certainty, not the accuracy of genetic selection.

12. Permutation City ; Greg Egan; novel, 1994

Mechanism: if minds became executable, computing resources would affect continuity, speed, copying and the meaning of survival. It offers unusually rigorous thought experiments about identity and the incentives surrounding artificial lives. Author’s publication history.

What to borrow: scarce processing time can become a class difference. A copy’s continuity and its relationship to its original may become emotionally and institutionally contested.

Limit: human uploading is not demonstrated. Egan explicitly says he does not take the book’s Dust Theory very seriously, explains technical shortcuts, and identifies ethical flaws he later saw in the premise. His self-critical FAQ is essential reading. This is a conditional thought experiment, not a forecast that physics permits substrate-free immortality.

Attractive works to use more cautiously

• Psycho-Pass (anime, 2012): useful for institutional deference to quantified judgments and for the separation of moral responsibility from an official score. Precise, universal measurement of criminal propensity from a person’s mental state is unsupported; it should be treated as the premise under examination. Production I.G’s original broadcast record.
• Serial Experiments Lain (anime, 1998): powerful for disembodied reputation, uncertain identity and intimate technological alienation. Its network metaphysics and collective consciousness are not a causal engineering model. Use it for subjective experience and audiovisual form, not to substantiate uploading. Original distributor’s series and episode descriptions; screenwriter’s record.
• Akira (film, 1988): study urban scale, youth abandonment, military secrecy and the visual communication of dangerous power. Telekinesis supplies the plot’s technological breakthrough; it is not supported future science. Its real-world resemblances cannot validate that mechanism. BFI film record.
• Minority Report (Steven Spielberg’s film, 2002): valuable for dramatizing the consequences of treating a fallible or contested forecast as grounds for coercion. Its precognition is supernatural; predictive policing is not a demonstrated replacement for precogs. Study the institution’s incentives and the integration of advertisements into a coherent visual world. Amblin’s official film page.
• Patlabor 2: The Movie (Mamoru Oshii, 1993): a particularly strong supplementary choice for fabricated security crises, ambiguous evidence and conflict among institutions. Its giant machines are less useful as predictions than its account of how a crisis changes legitimate authority. Studio synopsis and original release date.

These downgrades concern forecasting evidence, not artistic quality. A story can be a masterpiece and an unreliable literal model.

Six mechanisms worth projecting into an original book

Confidence labels concern direction over roughly 2026–2035, not exact dates, prevalence or a predicted world government.

1. Delegated economic power could expand before settled machine personhood. Present institutional work: documented; substantial adoption by 2035: medium confidence. Google’s Agent Payments Protocol announcement identifies authorization, authenticity and accountability as prerequisites for agent purchases. That is a protocol announcement and reference implementation, not measured deployment, transaction volume or proof of an autonomous economy. An original story could follow a disputed chain of delegated authority: the purchase was validly authorized, yet nobody intended its aggregate consequences. Watch insured transaction volume and responsibility for losses, not partnership counts.

2. Continuity could become a source of power. Scenario confidence: medium; magnitude and harm: unresolved. As a service accumulates personal context, leaving it can mean losing accumulated adaptation. Chiang, Egan and Iso illuminate different parts of that dependency. The MIT-led four-week study with 981 participants makes psychosocial effects empirically investigable; it cannot establish durable consumer lock-in or long-term harm. The switching-cost mechanism is a separate inference requiring direct evidence about portability, churn, willingness to pay and welfare after migration. Watch exportable memory, independent hosting, interoperability and the persistence of relationships after model changes. A portability right could matter as much as a privacy right.

3. Interfaces may increasingly observe the world alongside their users. Research direction: documented; widespread adoption and economic power by 2035: medium confidence. Meta’s Aria Gen 2 research combines egocentric sensing with contextual AI research. This establishes a research direction, not proof of a deployed advertising regime. Combine Feed and Dennou Coil: the useful assistant observes enough to become a powerful gatekeeper of attention. Counterforce: local processing, battery limits, conspicuous cameras, consent and user rejection. The consequence depends on governance and business models, not sensors alone.

4. Automation may reorganize hidden human work before eliminating it. Sector-dependent scenario: medium confidence. The π0.5 paper develops generalization across robots and unfamiliar environments using heterogeneous data. It demonstrates technical progress; it does not measure employment, paid intervention time or economic substitution. Sleep Dealer suggests a credible transitional world of remote exception handlers, demonstrators and local repairers. Treat this as an inference: some jobs disappear, some expand, and some become fragmented oversight. Watch paid human minutes per completed physical task and intervention frequency.

5. Experimental verification becomes a strategic bottleneck. Direction: high. Automated research is real, but interpretation remains consequential. The A-Lab paper’s January 2026 correction and updated article report 36 compounds from 57 targets over 17 days; the updated methods describe manual Rietveld refinement of the original X-ray diffraction data to check phase identification, including comparisons against known alternative phases. That reanalysis was external validation and did not guide the autonomous decision loop; it was not a newly repeated independent synthesis campaign. This is precisely why discovery headlines should be audited. A Lem-like institution could possess superior experiments but uncertain interpretations. Watch replicated performance, characterization quality and manufacturability rather than generated hypotheses.

6. Behavior becomes impressive before agreement on minds. Capability direction: high; consciousness conclusion: unresolved. Contemporary indicator-based consciousness research offers a framework, not a definitive detector. Meanwhile, functional mouse connectomics maps structure and recorded activity; it does not upload an organism. Blindsight and Permutation City are valuable as opposing pressures. A credible story can leave personhood disputed while making the systems socially indispensable. Do not resolve the scientific question by giving a machine one moving monologue.

Audit after cross-review

Confidence in a present research direction must not become confidence in mass adoption, coercion or durable market power. The numbered forecasts above were revised accordingly: consumer continuity, wearable observation and hidden-work reorganization are medium-confidence scenarios, not demonstrated outcomes. A four-week psychosocial experiment cannot establish lifelong switching costs; a robotics paper cannot establish labor-market effects; a payment-protocol announcement cannot establish transaction volume. Agreement among research agents is not independent evidence.

The A-Lab claim was checked against the updated original article and correction metadata. The relevant count is 36 compounds from 57 targets, not the original widely circulated 41/58. External manual validation means reanalyzing original diffraction measurements, not independently reproducing all syntheses. The paper supports bounded experimental autonomy while documenting an important interpretive limit.

Adversarial checks for the manuscript

The 2025 personalized GPT-4 persuasion experiment explicitly cautions that laboratory persuasiveness does not establish real-world effectiveness. A separate policy-persuasion study reports small average shifts. Therefore write influence as conditional and cumulative, with failed persuasion and resistance; do not substitute a magical command phrase for sociology.

For every technology, specify its operating cost, failure condition, maintenance organization and institutional owner. For every secret, specify who knows it, who is mistaken, who benefits from exaggeration and which physical trace would reveal it. For every major forecast, write a plausible counterfactual where regulation, cheap alternatives, public rejection or engineering constraints prevent it. For every impressive demonstration, distinguish one successful attempt from deployment over millions of variable situations.

A particularly strong synthesis for an original novel is Chiang’s dependence + Rivera’s labor geography + Lem’s disputed knowledge + Iso’s everyday infrastructure + Itoh’s benevolent coercion. These are compositional principles, not a plot to copy. They can support an ambitious future whose drama comes from understandable incentives even when its technologies are unfamiliar.

Access note: this research used author-hosted texts, official sources, authorized excerpts, library catalogues and open research. Blindsight is legally available in full from its author; Accelerando has authorized free editions. No shadow-library downloads were needed, and no inaccessible text is presented as having been read in full.

Fiction as a laboratory for society, biology and the material economy

Research date: 28 September 2026. Independent society-domain contribution for the future-fiction investigation. Selection is deliberately mechanism-first: a work earns its place by making a useful causal model, exposing an overlooked constraint, or furnishing a serious counter-scenario. It is not a ranking of which author “predicted the most.” Plot and publication checks use authors, estates, publishers and official distributors; factual tests use institutional data or original research. No unauthorized copies were acquired. The analysis contains spoilers.

Strongest conclusion: technical capability can grow while access to housing, insurance, care, reproductive services, reliable electricity and education remains unequal. Greater administrative restriction is one possible political outcome, not an empirical forecast established by these sources. Public provision, competition, regulation and cooperative institutions could instead widen practical access. These competing responses to persistent allocation problems are better grounded than a single worldwide collapse, total infertility, or inevitable technological unemployment. Fiction is especially valuable for examining how people normalize institutions; and how they change them.

Twelve works worth studying closely

1. Octavia E. Butler ; Parable of the Sower (1993), with Parable of the Talents (1998)

Useful mechanism: public services deteriorate unevenly; people seek protection in compounds, employer dependency and new communities. Earthseed makes meaning, obligation and collective purpose part of survival. This is more useful than treating fire, an authoritarian slogan or the calendar year as a prediction score. The estate explicitly describes Butler’s project as an exploration of problems and possible responses. Estate’s series account

Evidence and scenario: Treasury’s analysis of 2018–2022 insurance data found materially worse availability and cost in climate-exposed areas. Insurance is a mechanism connecting physical risk to housing finance and household mobility. A plausible 2030–2050 story therefore begins with an insurance renewal, utility shutoff or lost job-linked benefit before it reaches barricades. Treasury report

Limit: the novels are warnings, not an empirically established timetable of American collapse. Their collective institution-building deserves as much attention as their brutality.

2. Kim Stanley Robinson ; The Ministry for the Future (2020)

Useful mechanism: decarbonization requires changes in institutions, finance, incentives and legitimacy as well as inventions. Its many viewpoints keep engineers, bankers, victims and administrators in the same causal system. Publisher and date

Evidence and scenario: the UN adopted a Declaration on Future Generations in 2024, and central banks already use climate scenarios. These are institutional analogues, not evidence of a ministry with the novel’s powers. The IEA’s World Energy Investment 2026, published 28 May 2026, tracks substantial continuing investment in clean technologies and electricity infrastructure. An adaptive 2030–2050 future with immense conflict over implementation is more plausible than assuming the transition either succeeds smoothly or never occurs. UN declaration, NGFS, IEA’s dated report edition

Limit: the carbon currency, political coordination and geopolitical sequence are scenario inventions. A crisis can also entrench incumbents or fracture cooperation.

3. Paolo Bacigalupi ; The Water Knife (2015)

Useful mechanism: water scarcity becomes a contest over legal priority, infrastructure and enforceable documents. The resource is physical; the allocation mechanism is political and financial. Publisher

Evidence and scenario: Reclamation’s post-2026 Colorado River process establishes that allocation rules are actively contested. Its 2025 extension of 18 conservation agreements also demonstrates bargaining and compensated conservation. For a grounded 2035 narrative, follow agricultural transfers, tribal rights, municipal borrowing, reuse systems and who pays for conservation. Post-2026 process, Conservation agreements

Limit: assassinations and interstate breakdown are thriller escalations. Scarcity can produce cooperation and efficiency; it does not mechanically produce war.

4. Margaret Atwood ; Oryx and Crake (2003)

Useful mechanism: commercial biology, security enclaves, manipulated demand and private scientific authority become interdependent. Its strongest lesson is that a laboratory belongs to an institution with incentives and a distribution system. Publisher, Publisher reading guide

Evidence and scenario: gene-edited pig kidneys have entered a clinical trial, but organ rejection and other medical complications remain substantial. The novel’s biological supply-chain imagination is useful; engineered replacement people and neatly executed species-wide destruction are not established capabilities. NYU clinical-trial account

2030–2050 implication: biologic manufacturing, proprietary cell lines, trial access, reimbursement and biosecurity oversight can determine power. Write the mundane intermediaries that turn a scientific result into a privilege.

5. Kazuo Ishiguro ; Never Let Me Go (2005)

Useful mechanism: institutions can teach people to accept profound exploitation through kindness, euphemism, limited information and ordinary aspiration. Its emotional precision is more transferable than its transplant technology. Faber, First-edition record

Evidence and scenario: xenotransplant research offers a competing technical route to increasing organ supply, undermining any claim that raising human clones is the likely medical solution. The book instead models how beneficiaries compartmentalize the suffering that supports their lives. NYU research

2030–2050 implication: a credible story could concern care workers, trial subjects or people whose biological material becomes valuable under consent terms they did not understand. Those are proposed story mechanisms, not claims that these groups currently experience the novel’s regime. Preserve the distinction between a moral thought experiment and a technical forecast.

6. Alfonso Cuarón ; Children of Men (film, 2006)

Useful mechanism: loss of confidence in a common future transforms borders, solidarity, policing and personal motivation. The film sets its universal-infertility premise in 2027. Official distributor

Evidence and correction: the UN’s 2024 central projection still expects global population to grow to roughly 10.3 billion in the mid-2080s. Lower birth rates are not evidence of universal biological sterility. UN population projection

2030–2050 implication: a more realistic demographic conflict combines aging rich societies, different regional age structures, migration demand, contested parenthood and unequal access to care. Use the film for atmosphere, political psychology and the treatment of outsiders; reject its literal reproductive premise as a baseline forecast. It is a useful counterexample to confusing evocative cinema with biological evidence.

7. Hao Jingfang ; “Folding Beijing” (English publication, 2015)

Useful mechanism: elites monopolize desirable time and space; the excluded remain necessary to the system. A parent’s wish to fund a child’s education reveals the entire political economy. The folding city is a metaphorical device, not engineering foresight. The story is legally available in full, and Hao explicitly situates it within her research on inequality. Authorized text, Author’s essay

Evidence and scenario: OECD research documents housing access becoming less equal across generations. The 2030–2050 analogue is segmentation through commuting, school admission, work schedules, housing tenure and priority access; not moving buildings. OECD wealth study

Craft value: show the privileged city and the service city occupying the same geography at different hours. This connects abstract class structure to an immediately filmable experience.

8. Chen Qiufan ; Waste Tide (2013; English 2019)

Useful mechanism: technological consumption exports physical costs to workers and places that disappear from affluent users’ narratives. Local clans, migrant labor, foreign capital and ecological damage interact rather than collapsing into one all-powerful corporation. The author’s own account anchors the setting in the geography of his childhood. Author, Publisher and authorized excerpt

Evidence and scenario: the Global E-waste Monitor reports 62 million tonnes generated in 2022, with 22.3% documented as formally collected and recycled. This directly supports the material externality mechanism; it does not validate the novel’s speculative human-machine transformations. ITU/UNITAR report

2030–2050 implication: follow servers, batteries, sensors and robots through mining, replacement, repair and disposal. An advanced future has waste workers, chemical exposures and procurement fraud as well as inventors.

9. Makoto Yukimura / Goro Taniguchi ; Planetes (manga; anime 2003–2004)

Useful mechanism: expansion into space creates routine maintenance, dangerous work, hierarchy and conflicts over who receives the benefits. The anime’s orbital-cleanup crew has wages and ambitions, not just a mission patch. Official distributor’s synopsis and dates

Evidence and scenario: ESA’s 2026 space-environment report documents the growing operational problem alongside improving debris-mitigation compliance. ESA report

Limit: human crews picking up debris should not be assumed economically optimal. The near-term analogue is robotic servicing, tracking, collision avoidance, insurance and responsibility for disposal. Its 2075 routine human spaceflight is a conditional later-century scenario. The durable forecast is the emergence of maintenance institutions around a valuable shared environment.

10. Cory Doctorow ; Walkaway (2017)

Useful mechanism: exit, open designs, shared production and mutual aid compete with ownership and enclosure. The book tests whether abundance changes bargaining power, rather than assuming better gadgets automatically benefit everyone. Author and date, Publisher

Evidence and scenario: Ostrom’s empirical work establishes that communities can govern shared resources under identifiable institutional arrangements. That supports taking cooperative alternatives seriously; it does not prove self-sufficient maker economies or mind uploading. Ostrom’s primary lecture

2030–2050 implication: cooperative energy, repair, knowledge and some manufacturing could coexist with market and state systems. The hardest constraints are semiconductor production, medicines, land, feedstocks, logistics and coercion. A convincing optimistic society must solve those dependencies rather than simply print its way out of them.

11. Ursula K. Le Guin ; The Dispossessed: An Ambiguous Utopia (1974)

Useful mechanism: formal equality does not eliminate prestige, conformity, gatekeeping or career incentives. Conversely, scarcity does not make competitive private ownership the only imaginable arrangement. Shevek’s scientific work is inseparable from who may communicate it and who expects to possess it. Estate synopsis and publication history

Evidence and scenario: Ostrom provides empirical support for the feasibility of some commons institutions; Le Guin supplies the psychological and political stress test. Neither establishes that a planetary anarchist society will exist. Ostrom research summary

2030–2050 implication: compare research cooperatives, public laboratories, commercial labs and closed national projects. Let all have accomplishments and failure modes. This is essential protection against writing either corporate dystopia or communal utopia as a foregone conclusion. Faster-than-light communication remains a fictional premise.

12. Hitoshi Ashinano ; Yokohama Kaidashi Kikou (manga beginning in the 1990s; anime adaptations)

Useful mechanism: gradual contraction can contain hospitality, beauty, attachment, grief and adaptation. The android-run café gives civilizational change a local rhythm. Official English publisher

Evidence and scenario: aging and regional population decline are compatible with the UN’s growing global total. Sea-level change also proceeds unevenly and over different timescales; IPCC estimates do not support using dramatically submerged cities as the default 2030–2050 coastline. UN demographic data, IPCC sea-level assessment

Craft value: portray a shrinking town still operating, an inherited machine maintained beyond its design life, or a community deciding which road to abandon. This is an adaptive emotional counterweight to apocalypse, not an endorsement of inevitable human extinction.

Five checks against seductive but weak forecasts

1. Scarcity can change location rather than disappear or become universal. IEA’s 2026 report says low-emissions sources reached 43% of electricity generation in 2025. That weakens a simple return-to-muscle-power trajectory such as The Windup Girl’s energy setting. Grid access, storage, finance and reliability can remain scarce while generating equipment gets cheaper. Its seed-IP politics remain useful: USDA documents substantial seed-sector concentration. IEA, Windup publisher, USDA

2. A measurement need not be accurate to acquire social power. Gattaca is a better model of discrimination justified by genetics than of reliably specified human potential. ASRM’s December 2025 assessment finds polygenic embryo screening not ready for clinical use. A 2035 caste system based on overtrusted scores can be plausible even when biological determinism is false. Sony, ASRM

3. Reproductive technology will meet legal, developmental and family constraints. HFEA says laboratory-derived gametes may expand options, but treatment viability remains uncertain and their clinical use is prohibited under the UK framework described in its January 2025 statement. “Designer families soon” outruns this evidence. HFEA

4. Exposure is not replacement. ILO’s 2025 estimate places one-quarter of global employment in occupations with some generative-AI exposure. This is a task-based exposure model, not an observed job-loss rate; the authors consider transformation more likely than full replacement. A stronger story includes new supervision work, degraded entry-level ladders, labor bargaining and changed job quality alongside displacement. ILO

5. Neither automatic cooperation nor inevitable savagery is a good default. Water agreements, commons research and existing public services establish real adaptive capacity. They do not guarantee success under larger shocks. Keep disaster, reform, capture and recovery as competing branches, and identify which institutional variable changes the outcome.

Scenario synthesis for a book

Most defensible 2030–2050 direction, high confidence in the broad pressures: more electrification and digital capability coexist with aging, climate adaptation, expensive urban access, and fights over the distribution of security and care. Confidence is lower in exactly which countries or firms manage those pressures best. A technologically advanced city can have excellent cancer therapy and failing drainage; cheap machine-generated knowledge and inaccessible housing; abundance in one district and long queues in another.

Conditional constructive branch: shared infrastructure, reliable public administration, labor adaptation and commons institutions distribute more of the gains. Use Ministry, Walkaway and The Dispossessed to make cooperation demanding, internally contested and therefore dramatically credible.

Conditional extractive branch: benefits are bundled into employer, insurer or citizenship packages; leaving becomes expensive; excluded populations perform maintenance and absorb waste. Combine Parable, Folding Beijing, Waste Tide and Never Let Me Go. The “hidden network” is often discoverable in contracts, ownership, reimbursement rules, professional standards and subcontracting chains. Evidence of shared incentives is not evidence of a secret unified command.

Longer-horizon speculation: the biological redesign of humanity, portable memories, conscious uploads and routine interplanetary economies should be marked as major additional assumptions. They can make excellent fiction. They should not silently inherit the confidence of well-evidenced near-term developments.

For original worldbuilding, join one emerging capability to one slow institution and one intimate need: a family contesting an insurer’s reproductive score; a municipal engineer rationing reliable power; a repair cooperative dependent on a hostile chip supplier; a migrant technician maintaining a city they cannot legally inhabit. These are original prompts inferred from the combined evidence, not predictions or adaptations of the cited plots.

Adversarial cross-audit

The following challenges result from reading the technology and world-order strands. Agreement among the contributors is not additional empirical evidence.

1. A protocol announcement does not establish widespread delegated economic power. The technology strand's AP2 evidence supports work on authorization infrastructure. It does not establish adoption, transaction volumes, insurability or profitable end-to-end autonomy. Separate high confidence that delegation mechanisms are being built from lower confidence about the scale of autonomous economic activity by 2035. A counter-scenario is extensive agent assistance with humans retaining approval at consequential steps.

2. The chatbot study does not test commercial lock-in. The cited four-week psychosocial experiment is relevant to attachment and behavior; it does not establish switching costs, portability failures or market concentration. Continuity becoming a major source of corporate power remains an inference. Test it with longitudinal retention, cross-provider migration, export functionality and the cost of reconstructing useful context. Multi-provider use and interoperable memory could weaken this mechanism substantially.

3. Spending is not enough to infer a change in the character of security. SIPRI's aggregate military series establishes resource commitment. It cannot by itself demonstrate that maintenance, repair or network resilience is becoming more strategically important than before. That forecast needs procurement composition, readiness, repair times, spare-parts availability and operational outcomes. The alternative is that cheap expendable systems reduce the value of maintaining particular expensive platforms while increasing the importance of rapid replacement production.

4. Climate vulnerability does not determine a political regime. Treasury's geographic insurance associations do not prove that climate alone causes nonrenewal, still less that corporate feudalism follows. Construction costs, regulation, reinsurance, public insurance and adaptation also matter. Lower birth rates likewise do not establish biological infertility; useful genetic associations do not establish individual destiny. Keep these distinctions explicit in the final synthesis. Faster emissions reduction would improve longer-term outcomes but would not immediately falsify near-term adaptation pressure because infrastructure exposure and committed warming persist.

5. The corpus is selected for causal insight, not demonstrated forecasting accuracy. Publisher descriptions and authorized excerpts verify premises but do not amount to new full-text readings of every work. The three strands overrepresent institutional conflict, so their thematic convergence cannot independently prove that conflict is the most likely future. Include constructive branches and record each work's misses. For a true prediction-accuracy claim, define the candidate corpus, extract predictions made before the outcome, include failures, and compare against contemporary expert baselines. This project has not performed that quantitative exercise.

World order, institutional fiction, and falsifiable futures

Research cutoff: 28 September 2026. Prepared as a specialist contribution to a broader fiction-and-forecasting investigation. Public sources only. The forecasts below are the researcher's judgments; sources establish starting conditions, not the truth of a future scenario. Confidence labels are qualitative and are not statistically calibrated probabilities.

Principal finding

The strongest baseline for 2030–2050 is overlapping, contested infrastructure blocs. Countries can depend on American financial and software networks, Chinese industrial supply chains, Gulf capital, and local political coalitions simultaneously. The map changes according to whether one traces semiconductor tooling, power, banking, military protection, or food. A single ranking of “the next superpower” conceals these differences.

The most useful science fiction accordingly models permissions, maintenance, bargaining, legitimacy, and unequal access. Its gadgets can be wrong while its political mechanism remains useful. No novel has demonstrated an ability to accurately forecast an entire future. Resemblances noticed after the event are vulnerable to cherry-picking.

What the money and physical evidence actually establish

Sources: TSMC management report; IEA update; Microsoft–G42 disclosure; Aligned acquisition completion; SIPRI expenditure release.

The unusually useful network edges are equity ownership, a board seat, a power-purchase agreement, an export license, a secured loan, a government procurement contract, and a cloud dependence. Each conveys different powers. A shared investor alone is weak evidence of operational coordination. Asset managers frequently manage other people's capital; assets under management are not their freely spendable fortune.

A particularly valuable source reaches beyond press releases: the FTC's January 2025 study used compelled nonpublic partnership documents. It identifies cloud spending commitments, varying control and exclusivity rights, information sharing and potential switching costs in three partnerships. The observed documents run through September 2024; the report neither proves illegal conduct nor establishes every contract's current terms. This is strong evidence for contractual mechanisms and weak evidence for claims of a single coordinated agenda. FTC report

For a fictional conspiracy to feel credible, give it an enforceable mechanism: a debt covenant, proprietary repair key, required safety certification, monopoly input, or emergency procurement exemption. “They all attend the same conference” is insufficient.

Seven mechanisms, forecasts, and ways they could fail

1. AI sovereignty becomes a negotiation over physical infrastructure

Evidence: In the 2026 IEA update, electricity, grid connections, chips, equipment supply and finance constrain deployment. Its 2030 central electricity projection is around 3% of global electricity, so the bottleneck is often local concentration rather than global exhaustion. IEA

2030–2035 forecast ; high confidence in the constraint; medium in the political outcome: Governments seeking frontier AI capacity will bargain across several systems at once: access to accelerators, sufficiently reliable electricity, financing, data jurisdiction and a trusted operating partner. “We own the building” will frequently coexist with dependence on a foreign hardware or software stack. The Gulf can convert capital and infrastructure into bargaining power, while remaining exposed to suppliers and authorization regimes. This is an inference, not a claim that all Gulf states have a common strategy.

2035–2050: An important division may be between countries able to maintain and replace their computing infrastructure and countries able only to rent its services. Small efficient models and open systems can weaken this division. Owning a frontier training cluster will not be necessary for every economically useful deployment.

Disconfirming evidence: Frontier training and broad deployment become inexpensive on commodity hardware; open software substitutes for proprietary stacks; firm power and chip capacity develop large sustained surpluses; numerous independent operators reach comparable capability without state-scale financing.

Fiction use: A city negotiates political concessions for guaranteed compute and cooling during heat emergencies. Its ruler's vulnerability is a maintenance contract, not a more intelligent enemy chatbot.

2. Fragmentation produces brokers as well as blocs

Evidence: A 2025 IMF working paper separates production relocation from simple transshipment. In its six-country Asian sample, Vietnam shows evidence of increased domestic value added in strategic exports to the US; evidence for the other five is less clear. This cautions against treating every connector economy as tariff evasion. The paper is authors' research, not an official IMF forecast. IMF working paper

2030–2035 forecast ; medium-high: Selective decoupling and continued interdependence coexist. Countries able to provide credible local production, political reliability and access to several markets gain leverage. Their gains are conditional: stricter rules of origin, sanctions enforcement or war can eliminate their position.

2035–2050: India, Indonesia, Vietnam, Turkey, Gulf states and others should be written as actors bargaining across relationships, not an interchangeable “Global South.” Their incentives and domestic constraints differ. This is a scenario design inference, not a prediction that all become great powers.

Disconfirming evidence: Broad and enforceable prohibitions force comprehensive alignment; major cross-bloc trade and investment cease; or a durable political settlement reverses discriminatory restrictions. Track domestic value added, physical factories and customs data rather than export growth alone.

Fiction use: A supposedly neutral port maintains incompatible compliance systems; choosing which cargo to certify becomes a constitutional crisis.

3. Industrial power rests on processing and substitution, not simply owning deposits

Evidence: The IEA's 2026 minerals outlook emphasizes refining concentration, difficult substitutes and materials produced as by-products. Its outlook identifies China as the top refiner, above 90% supply in the particular gallium, graphite, manganese and rare-earth categories discussed. Categories and processing stages must not be conflated with mined output. IEA mineral outlook

2030–2035 forecast ; high for processing importance; medium for persistence of today's leaders: Refining, chemical expertise, qualification and industrial customers continue to determine who captures mineral value. Resource-rich countries can improve their leverage through reliable processing and downstream demand, but a mineral discovery by itself is weak evidence of future geopolitical dominance.

2035–2050: Recycling, alternative chemistries and redesign can transfer leverage between actors. “Lithium is the new oil” is an especially poor permanent worldbuilding rule: different materials have different substitution and reuse pathways.

Disconfirming evidence: Competing refineries achieve sustained commercial output with qualified customers; material intensity falls substantially; a substitute technology scales rapidly. Announced capacity and a laboratory chemistry are not enough.

Fiction use: An industrial federation survives an embargo because its repair culture recovers scarce components. Its power is in accumulated production knowledge and scrap inventories.

4. Financial networks change more slowly than payment interfaces

Evidence: The IMF reports the dollar at 57.13% of foreign-exchange reserves in 2026 Q1 and the renminbi at 1.99%; valuation movements explain part of changes. Separately, BIS reports the dollar on one side of 89.2% of FX trades in April 2025. Currency shares of FX trades sum to 200%, because each trade has two currencies. Reserves and transaction turnover are distinct metrics. IMF COFER, BIS 2025 survey

2030–2035 forecast ; medium-high qualitative confidence after adversarial review: The dollar remains the largest international reserve currency, even as bilateral settlement and payment platforms diversify. This is a narrower claim than stable reserve share or unchanged US financial influence. A new payment rail is not itself a new reserve asset, safe-asset market or lender of last resort.

2035–2050 ; medium to low: A more plural monetary order is plausible, but its form depends on financial openness, credibility, debt-market depth and security arrangements. It cannot be inferred from blockchain pilots or an announced BRICS currency.

Disconfirming evidence: Several years of large valuation-adjusted reserve shifts; deep and liquid alternative safe-asset markets; sustained non-dollar invoicing outside politically restricted transactions; credible crisis liquidity support at scale. A payment system's transaction count alone is inadequate.

Fiction use: A state's sovereign digital currency works technically, but importers still need foreign collateral and insurers. The crisis arrives at settlement and refinancing.

5. Demographic divergence makes migration, care and skills central to power

Evidence: UN projections put sub-Saharan Africa at about 2.2 billion people in 2054, a 79% increase from the report's starting point. These are conditional projections; end-century fertility uncertainty is much larger. The World Bank calculates that India's aspiration of high-income status by 2047 would require average growth of 7.8% over roughly two decades. That is a required pathway, not a forecast. UN demographic policy brief, World Bank India analysis

2030–2050 forecast ; high on divergent age structures; medium on consequences: Aging societies bargain over migration, retirement, care and automation. Young populations confer an opportunity conditional on education, health, housing, reliable power and employment. A large population is neither automatic national strength nor an automatic instability trap.

2050–2100 ; conditional: African societies become much more consequential in population terms under current UN trajectories. Which states, cities or institutions convert this into productive and political power is substantially less predictable. Forecast urban systems and corridors individually; Africa is not a unitary actor.

Disconfirming evidence: Persistent fertility departures from UN trajectories; transformative migration agreements; unexpectedly fast productivity/care automation; institutional changes that break the assumed link between age, work and fiscal burden.

Fiction use: A country imports caregivers and engineers while denying them durable political membership, creating a conflict between demographic need and national identity.

6. Climate adaptation becomes a contest over protected assets and fiscal capacity

Evidence: UNEP estimates developing-country adaptation needs in 2035 at US$310 billion annually from modeled costs, or US$365 billion from extrapolated national plans; international public adaptation finance was US$26 billion in 2023. These are different measures and years, not a contemporaneous complete global spending gap. World Bank Groundswell's widely quoted “216 million” is an upper scenario estimate for internal climate migration across six regions by 2050, not a prediction of cross-border refugees. UNEP 2025 report, World Bank Groundswell

2030–2050 forecast ; high on unequal pressure, medium on the political form: Governments and financiers decide which neighborhoods, ports, farms and power facilities receive protection, cooling or relocation funding. A plausible consequence is unequal access to habitable and insurable places. Whether that becomes authoritarian exclusion, public investment or negotiated relocation is a political variable.

2050–2100: Sea-level and heat pressures persist beyond individual investment cycles. IPCC AR6 gives scenario-dependent likely global mean sea-level rise by 2100, relative to 1995–2014, from 0.28–0.55 m under very low emissions to 0.63–1.01 m under very high emissions. These are not local flood depths or a probability distribution across emissions scenarios. IPCC physical-science summary

Disconfirming evidence for unequal outcomes: Sustained adaptation delivery closes vulnerability gaps; equitable retreat and land policies prevent exclusion. Faster emissions reductions change longer-term pressure but do not immediately remove near-term adaptation needs or reverse committed sea-level rise. Compare actual protected populations and delivered infrastructure, not the number of national adaptation plans.

Fiction use: A functioning coastal megacity selectively defends economically valuable districts while quietly withdrawing mortgage guarantees elsewhere. The future is governed retreat, not uniformly submerged skylines.

7. Security is increasingly a problem of maintaining complex systems under stress

Evidence: SIPRI's rising expenditure series supports renewed resource commitment to security, but not a specific technological winner. ESA's September 2026 environment report identifies continued orbital crowding, collision risk and the need for active debris removal; its baseline includes activities through 2025 rather than every announced constellation. SIPRI, ESA

2030–2035 scenario ; medium confidence: Repair capacity, redundant communications, spare components, resilient navigation, logistics and institutional authority could become more strategically valuable where essential services depend on fragile interconnections. SIPRI's aggregate expenditure and ESA's orbital risk findings do not establish that this is a global procurement trend. Specific awards, delivered capabilities and operational results are needed to test it. A capable autonomous system can still fail because of power, parts, authentication or a dispute over who may order it to act.

2035–2050: Security bargaining includes private infrastructure operators and public authorities. This does not imply corporations replace states; the ability to license, tax, requisition and physically protect infrastructure remains central.

Countercase and disconfirming evidence: Cheap expendable systems and rapid replacement capacity can outperform investment in repairing or hardening individual assets. Reliable inexpensive redundancy can reduce outage consequences; operators may become independent of proprietary maintenance; orbital debris and liability may become effectively governed. Track specific procurement awards and delivered performance, including total replacement-versus-repair costs, rather than inferring preferences from aggregate defense spending.

Fiction use: A political thriller turns on the authority to recognize an authenticated emergency order. All technical systems function as designed; the institutions disagree on legitimacy.

Regional implications: scenario judgments, not league tables

This table is a comparative hypothesis based on the mechanisms above, not a separately measured forecast. A useful corrective for claims of inevitable European decline is that the EU itself has mapped investment and coordination problems and proposed responses. Proposals are not implementation. European Commission's Draghi report portal

Competing world orders for an author's scenario set

1. Managed rivalry with dense interdependence ; strongest baseline judgment through the 2030s. Restricted strategic sectors coexist with large civilian flows. Middle powers bargain; many firms operate several incompatible systems. This follows the fragmentation and network evidence without presuming complete separation.
2. Hard security blocs ; serious downside scenario. A major war or sanctions crisis forces broad alignment, duplicative production and sharply higher costs. Its likelihood cannot be responsibly inferred from military expenditure alone.
3. Competitive abundance ; plausible counter-scenario. Cheap clean electricity, efficient AI and widespread industrial learning diffuse capability. Competition continues, but today's chokepoints lose some importance. This is the necessary antidote to assuming every current shortage lasts forever.
4. Uneven archipelago ; plausible political outcome across several scenarios. Highly functional urban and industrial regions coexist with poorly protected areas within the same country. National GDP growth can mask fragmentation in lived security.

For 2050–2100, these should remain branches rather than a single predicted map. Exact regimes, borders, wars and hegemonic succession have low predictability at that horizon. Demography and physical climate processes support stronger constraints than they support specific political stories.

Eight fiction works to study for these mechanisms

Ratings here concern usefulness for institutional imagination, not a measured probability that the plot happens. Primary publisher or studio descriptions verify the works and central premises; the critical judgments are this researcher's reading. A synopsis was not treated as a substitute for fresh full-text analysis.

For this particular geopolitical assignment, Patlabor 2, 2nd GIG, The Ministry for the Future and The Dispossessed form the highest-yield core. They train attention on authority, institutions and competing obligations. Planetes is the corrective that makes infrastructure feel worked on and paid for. None is an oracle.

Additional retrospective audit: Stand on Zanzibar ; John Brunner (1968). Its media collage, corporate power and differentiated forms of knowledge remain valuable writing methods. Its around-2010 population scale was close to the UN's contemporaneous seven-billion milestone in 2011; that is not a failed numerical forecast. However, extrapolating visible population growth does not validate the associated social scenario, technological dates or other predictions. Assess the whole package, not a viral list of isolated matches. This contribution did not perform a new full-text forecast census. Publisher and authorized excerpt, UN's 2011 milestone announcement

Adversarial second pass: corrections and decisive tests

Dollar continuity: The original high label was too strong for an unconditional nine-year political forecast. It is now medium-high for the precise outcome “largest reserve currency through 2035,” with materially lower confidence beyond that. Starting share and network effects do not exclude a US institutional crisis, restrictions on convertibility, war or a coordinated alternative. Decisive indicators are persistent valuation-adjusted reallocations and alternative crisis-liquidity capacity, not a one-quarter reserve change. This forecast says nothing about investment returns or the dollar's exchange rate.

China's industrial leverage: A current statistical anchor strengthens the starting-condition claim. UNIDO's June 2026 report ranks China first among manufacturers and reports Q1 production growth of 1.8% quarter-on-quarter and 5.6% year-on-year. These are harmonized industrial production indices, with official statistics supplemented by estimates; growth is not a profit measure or a forecast of geopolitical victory. The same report records 2.2% quarterly growth for Asia-Pacific excluding China, demonstrating that China is not the sole regional source of dynamism. A decisive long-run test is whether high output translates into durable profits, innovation, domestic welfare and coercive leverage when customers diversify. UNIDO Q1 2026 report

Cheap/open AI counterforce: It weakens dependency most directly for useful deployment and inference. It does not by itself remove frontier training, chip fabrication, electricity or proprietary data dependencies. Efficiency can also expand demand. The two claims must coexist: widely distributed useful AI and concentrated frontier infrastructure are compatible. The decisive tests are the share of valuable tasks runnable locally, switching costs, quality-adjusted total cost and how many organizations can train a competitive frontier system independently. Do not infer geopolitical democratization from downloadable model weights alone.

Money and energy audit: The IEA values were rechecked against the 2026 publication; distinguish its estimated historical electricity use from its projected 2030 result. Do not mix the IEA calendar labels with a company's fiscal-year capex without reconciling accounting periods and leases. TSMC's figure is the company's reported 2025 consolidated expenditure. The Aligned closing was independently rechecked on the acquired company's own site: completion, enterprise valuation, and an additional US$5 billion growth-capital commitment are separate facts. Its 6.4 GW figure combines operational and planned capacity, so it must not be presented as energized supply. Aligned's own closing release

One more counterweight: A country can lead one manufacturing stage and depend on foreign equipment at another. ASML's 2025 report records substantial sales and service activity, including stronger-than-expected DUV business in China. This supports the interdependence mechanism; it does not mean China has unconstrained access to every lithography technology. ASML 2025 financials

A concrete original premise produced by the evidence

Working scenario: a federated coastal corridor, 2042. Several cities finance cooling, flood protection and a major computing campus through an international consortium. Cheap automated research improves agriculture and materials, so society is visibly more capable than today's. However, the contract prioritizes industrial electricity during shortages; international insurers require an imported safety model; the model's license depends on geopolitical compliance. A neighboring population supplies workers but lacks equal residency rights. A storm exposes contradictory emergency mandates.

The protagonists could be an infrastructure auditor, a migrant maintenance engineer, a municipal finance official and a scientist whose open discovery threatens the consortium's contractual position. Every faction has a defensible objective. The contested resource is the power to decide which promise is broken first. The scenario combines forecasts without claiming that this exact city or crisis will exist.

Research standards and continuing audit

• Separate appropriation, commitment, financial close, construction, energized capacity and actual utilization. For money, record nominal versus real values and distinguish capex, equity value, enterprise value and assets under management.
• Trace network edges one by one. Document which party can terminate service, seize collateral, veto governance, deny a license or replace a supplier. Do not infer a hidden master plan from overlapping affiliations.
• Audit a forecast before it becomes a story assumption. Record date, mechanism, horizon, confidence, alternative explanation and disconfirming observation.
• Include failed technology predictions from each favored fiction work. Evaluate complete mechanisms rather than collecting isolated “hits.”
• Borrow local expertise for each setting: political economy, language, class, religion, housing, law and practical infrastructure. A non-Western location does not automatically create a non-Western viewpoint.
• For deeper follow-up, inspect securities filings, procurement awards, customs and value-added trade data, grid connection queues, environmental permits, local planning hearings, patent families and project lenders. These are research directions, not sources claimed to have been exhaustively inspected here.

Unresolved limitations: no exhaustive inspection of all relevant fiction; no access to classified systems, private contracts or unpublished ownership details; this contribution is a mechanism-based synthesis rather than a quantitative forecasting model. The strongest findings are present physical and institutional dependencies; end-century political predictions remain much weaker.

The technological frontier: evidence, restricted access, and strategic bets

Research cut-off: 28 September 2026. This is a selective investigation of high-information research, official programs, corporate commitments, and documented disclosures. I screened paper abstracts and inspected selected full methods, results, and limitation sections, alongside institutional records and firsthand accounts. It is not an exhaustive review of every paper, an independent reproduction of experiments, or access to classified systems. Publication dates, experiment dates, and future targets are distinguished below. Public documentation is much richer for some institutions and countries than others; that visibility difference cannot establish their relative secret capabilities.

The most consequential direction I identify is the automation of the discovery process itself: proposing designs, testing them against an external standard, retaining improvements, and using those improvements to conduct more research. Public evidence already shows pieces of this process in learning algorithms, computing infrastructure, cryptography, biology, and quantum control. The degree of autonomy, generality, and independent validation varies substantially.

The evidence does not establish a universal military or intelligence-service lead of 15 years. It does establish unequal access to models, data, infrastructure, and operational environments. Those advantages need separate analysis.

For this report, a research result means a result reported with a scientific method, not necessarily independently reproduced. External corroboration means another party confirms a particular outcome. A program or roadmap proves an activity or intention, not its success. An inference is my interpretation of the evidence. An authentic document can contain an unproven claim; a real company can have an unfulfilled roadmap.

1. There is direct evidence about the gap between internal and public AI.

METR assessed models shared by Anthropic, Google, Meta, and OpenAI during February–March 2026. Its May report estimated an average internal lead of roughly two months on a software-task time-horizon measure. The estimate is uncertain, partly because the benchmark was becoming saturated. The participating companies identified their supplied models as their internal frontier; METR did not independently inventory every system. Participants could redact disclosures or withdraw silently. The findings therefore constrain claims about these labs in that assessment window, not every classified program or September capabilities. METR, Frontier Risk Report

A genuine historical counterexample to complete public visibility is GCHQ's early public-key work, developed in the 1970s and disclosed in 1997. The delay before acknowledgement is not the same thing as the lead over independent civilian discovery. Likewise, the CIA's A-12 demonstrates that a specialized technology can be developed secretly; it does not establish a fixed lead across all technologies. GCHQ history, CIA A-12 history

The useful distinction is between invention lead, release delay, access restrictions, and deployment advantage. These are four different quantities. A government can achieve a large operational advantage using commercially familiar algorithms with unique intelligence collection and integration.

2. Restricted frontier AI exists, but restrictions need not imply a different underlying intelligence.

Anthropic's current Mythos page identifies Mythos 5.1 as restricted to vetted organizations. It also says the public Fable 5.1 uses the same underlying model, with additional safeguards that limit particular cyber and biology tasks. That is unusually direct evidence of capability-access differences. It is not evidence that the public model is universally years behind. Anthropic's current model and access description

Project Glasswing is a concrete restricted-access program for defensive security. Its expansion announcement describes institutional access and the growing bottleneck of reviewing, disclosing, and patching findings. Treat its aggregate discovery counts as vendor reports unless the individual findings have been validated. Glasswing expansion, 2 June 2026

An important external check comes from curl maintainer Daniel Stenberg. His team reduced five claimed security vulnerabilities in one Mythos report to one low-severity vulnerability, three false positives, and one ordinary bug. He also found useful bug reports. This is evidence of uneven performance and the need for expert triage, not a disproof of every stronger result on other projects. Stenberg's account and subsequent report disclosure

3. AI-assisted cryptanalysis is a more significant frontier than generic autonomous coding.

In July 2026 Anthropic reported AI-assisted advances against the HAWK signature candidate and reduced-round AES. NIST subsequently recorded that HAWK's developers withdrew it from consideration. This external institutional consequence makes HAWK a particularly strong example of AI contributing to consequential mathematical security research. NIST explicitly says its finalized post-quantum standards are unaffected. NIST's account

The AES result concerns seven rounds of AES-128, whose full version has ten rounds. The technical paper retains an enormous chosen-plaintext requirement and describes theoretical computational improvements; it does not demonstrate practical decryption of ordinary AES-protected traffic. Cryptanalysis of 7-Round AES via the Algebraic Structure of its S-box

The broader research question is whether models can discover a mathematical shortcut that experts missed, then support a checkable argument for it. CryptanalysisBench supplies 191 tasks across several primitive families, with known-broken, experimental, and production-strength targets distinguished. Success on its easier tiers must not be read as success on its hardest tier. CryptanalysisBench, July 2026

My inference: specialized, expensive search campaigns with mathematical or executable verification deserve more attention than a single chat response. The scarce resource is increasingly the combination of a good problem formulation, sufficient compute, and a reliable test.

4. Machines are beginning to improve learning rules, not merely produce more agents.

DiscoRL uses meta-learning over the experiences of a population of agents to discover the rule by which agents update their policies and predictions. Its authors report strong Atari results and transfer to other benchmark environments. This is a concrete instance of automatically designing part of the learning process. It is still a bounded experimental system; the paper does not establish unlimited recursive improvement or autonomous redesign of the entire research enterprise. Oh et al., Nature, 2025

AlphaEvolve provides a complementary approach: generate candidate programs, execute them, measure performance and correctness, and retain successful changes. Google reports using its outputs in TPU design and infrastructure optimization. The relevant advantage is access to internal workloads and hardware measurements, which outsiders cannot reconstruct from a model interface alone. AlphaEvolve infrastructure and research update, May 2026

A narrower paper reports automated optimization of fully homomorphic encryption kernels on TPUv5e. Within its experiment, TFHE bootstrap latency improved 2.5 times over a human-engineered baseline; CKKS operations improved by smaller factors. These are particular kernels and baselines, not a universal speedup for encrypted computing. Gorantala et al., May 2026

The frontier to monitor is a durable improvement in research productivity after accounting for search cost, failed trials, human checking, and transfer to new workloads. A demo that produces a clever algorithm is a weaker result than repeated improvements that survive production deployment.

5. Memory and learning during use are distinct technical frontiers.

DeepSeek-associated Engram research adds conditional lookup memory alongside conditional computation. It stores recurring local patterns in a large lookup structure so the network can spend more computation on other work. Its experiments compare equal-parameter, equal-compute baselines, including a 27-billion-parameter configuration. This is architectural memory for static patterns, not evidence that a deployed assistant automatically acquires permanent personal knowledge from every conversation. Conditional Memory via Scalable Lookup, revised July 2026

Nested Learning treats a model as interacting optimization processes operating at different update frequencies. Its Hope architecture explores self-modification and multiple memory timescales. The reported language-model comparisons include approximately 760-million- and 1.3-billion-parameter models. The scale matters: promising results at those sizes do not establish that catastrophic forgetting or reliable lifelong learning has been solved for frontier systems. Nested Learning, NeurIPS 2025 version and full paper

Coconut investigates reasoning through continuous hidden states rather than expressing every intermediate step in words. This is relevant to efficiency and to how much an observer can infer from a displayed explanation. Its existence does not establish that every closed model uses the method. Training Large Language Models to Reason in a Continuous Latent Space, revised August 2026

My assessment: persistent adaptation would be a major advance because it could turn experience into durable competence. The decisive tests are retention, transfer, resistance to bad updates, and predictable behavior after long use. More context storage alone does not satisfy those tests.

6. Formal verification can make some research loops unusually powerful.

AlphaProof combines reinforcement learning, proof search, and Lean's formal proof environment. For difficult problems it also uses test-time reinforcement learning over related problem variants. The paper explains the underlying system behind its 2024 competition results; the publication date should not be mistaken for the date of the original demonstration. Olympiad-level formal mathematical reasoning with reinforcement learning

The strategic idea is broader than competition mathematics: where correctness can be checked reliably, machines can generate many failed attempts without requiring a person to judge each one. This makes formal mathematics, selected software properties, circuit optimization, and some cryptographic research attractive early domains. My inference is that the quality and coverage of verifiers can matter as much as improvements to the generator.

The limitation is specification. A proof establishes a formal statement under its assumptions. It does not establish that the statement captures the intended real-world problem, that a simulator models reality, or that an optimized objective is socially desirable.

7. AI research is moving into biological experimentation, with human laboratory work still crucial.

Anthropic's 23 September report describes a search that identified array-associated reverse transcriptases with unusual repeat structures. The account reports roughly 950 agents, 21 hours, and 210 million tokens, followed by human laboratory work. The enzyme system's function remains unknown. This is an early discovery report, not a demonstrated new gene-editing treatment. Anthropic's biological discovery report

Biohub's April 2026 Virtual Biology Initiative commits $500 million over five years, including large-scale data generation and technologies for measuring and engineering biology. Mark Zuckerberg and Priscilla Chan identify AI-powered biology as their primary philanthropic direction. These are commitments to a research program, not evidence that a complete predictive human cell or a cure-all already exists. Virtual Biology Initiative, founders' statement

The deeper advantage is experimental feedback. Public literature is only one source of training material; measurements of what actually happens after an intervention can be much more valuable for learning causation. My inference is that institutions able to repeatedly design, run, and measure experiments can build advantages that will remain even if model access becomes broadly available.

8. World models and robotics aim to learn consequences, but physical reliability remains a barrier.

V-JEPA 2 investigates learning predictive representations from video and using a post-trained action-conditioned model for robot planning. Its authors report less than 62 hours of robot video for that post-training stage; this is not the total data used to develop the whole model. V-JEPA 2 paper

SIMA 2's work with Genie 3 explores agents acting and learning inside generated interactive environments. The important possibility is scalable generation of training situations. The unanswered issue is how faithfully useful skills and causal relationships transfer from generated worlds to unfamiliar physical conditions. SIMA 2 research account

Google's July 2026 Robotics 2 release provides a useful reality check alongside its demonstrations: some reported multifinger task success rates are only 32–44%. Its VLA and on-device models are described as available to early-access partners. Both capability gains and incomplete reliability are visible in the same primary source. Gemini Robotics 2 results

The next convincing evidence would be sustained task completion across unfamiliar settings, measured intervention frequency, recovery from failure, and operating cost. A polished video cannot answer those questions.

9. AI control of quantum hardware is already experimentally demonstrated.

A 2026 Nature paper uses error-detection events from quantum error correction as feedback for a reinforcement-learning controller. Rather than stopping everything to recalibrate, the controller adjusts parameters during computation. On Willow, the authors report 3.5-fold better logical stability against injected drift using control plus decoder steering. That qualifier matters: this is a particular controlled experiment. Reinforcement learning control of quantum error correction

This follows the earlier Willow demonstration of error correction below the surface-code threshold, where increasing code size improves logical protection. It does not by itself demonstrate a large fault-tolerant machine executing useful arbitrary algorithms. Quantum error correction below the surface code threshold

My inference is that AI may accelerate useful quantum computing through calibration, decoding, circuit design, and error management before quantum computers substantially accelerate general AI training. Those are different directions of influence and should not be conflated into a generic “quantum AI” claim.

10. Quantum resource estimates are improving faster than simplistic qubit-count headlines suggest.

The 2026 Pinnacle architecture paper estimates RSA-2048 factoring with fewer than 100,000 physical qubits under specified assumptions, including a physical error rate of 0.001, a one-microsecond code cycle, and fast classical reaction. Its detailed analysis associates that low-qubit configuration with approximately month-scale runtime. This is an architecture and resource estimate, not a completed factoring demonstration. Pinnacle, revised May 2026, especially Table 6

A separate 2026 whitepaper estimates resources for attacking 256-bit elliptic-curve cryptography, with configurations below 1,200 logical qubits or below 1,450 logical qubits and different gate budgets. Physical resources and runtime depend heavily on the architecture. Its revised version also records a correction to its zero-knowledge proof software, a reminder to inspect version histories. Babbush et al., revised April 2026

IBM's roadmap targets Starling in 2029, with 200 logical qubits and 100 million gates. DARPA's QBI asks whether industrially useful quantum computing can be achieved by 2033 and explicitly includes independent verification. Both are future targets. DARPA's separate selection of Microsoft and PsiQuantum for further validation confirms serious evaluation, not completed utility-scale machines. IBM roadmap, QBI, US2QC selection

To assess a claim, ask for the logical error rate, supported logical operations, sustainable circuit depth, cycle time, connectivity assumptions, and full cost against the best classical alternative. Physical qubit count alone is insufficient. Neither a quantum annealer nor a random-circuit sampling demonstration establishes a capability to run cryptographically relevant Shor computations.

11. The strongest Web3 signals involve settlement, verifiability, privacy, and payment infrastructure.

The BIS's May 2026 Project Agorá report describes a prototype combining tokenized commercial-bank deposits and central-bank reserves for atomic multicurrency settlement. The reported collaboration involved seven central banks and more than 40 regulated institutions. “Prototype” is essential: it is not evidence that a new global monetary system is already operational. Project Agorá report

BlackRock announced two additional tokenized money-market products in August 2026. This is a concrete institutional product direction. Tokenization represents financial claims through a new technical system; it does not automatically remove custodians, legal enforcement, or counterparty risks. BlackRock announcement

Ethereum's September protocol-priorities statement sets a December 2029 target for quantum resistance across execution, consensus, and data, alongside research on privacy, state, finality, and zkEVMs. It expressly describes the quantum planning assumption as aggressive and uncertain. A migration target is not a prediction of the date on which current cryptography will fail. Ethereum Foundation priorities

The x402 Foundation's July operational launch illustrates another direction: standardized payments within HTTP interactions, including payments by software agents. Its stated scope includes traditional cards as well as stablecoins. It therefore does not imply that every agent payment needs a cryptocurrency token. x402 Foundation announcement

The technical convergence I would watch is outsourced computation that can be checked or kept private, together with controlled machine spending. A zero-knowledge proof can attest that a specified computation was performed; it does not certify that the input was truthful or an AI conclusion is correct. Fully homomorphic encryption protects data during computation, but cost remains a material barrier. Buying a token is not the same as acquiring economic exposure to whichever technology succeeds.

12. Public military programs reveal specific frontier questions, not a single concealed master technology.

Quantum sensing and timing deserve separate attention from quantum computing. Their useful outputs may be navigation, measurement, or synchronization rather than faster arithmetic. Publicly documented engineering programs of this sort are better evidence than anonymous assertions that intelligence agencies already possess every futuristic capability.

13. Capital and institutions point toward control of infrastructure, experiments, and distribution.

The table is a strategic sample, not a ranking of net worth or an inventory of every investment. Amounts refer to different things; annual capex guidance, multi-year commitments, project targets, and transaction values; and should not be added together.

My inference from these commitments is that institutions are trying to control several inputs to discovery at once: compute and energy, domain-specific measurements, physical experimentation, and access to customers or government missions. This does not require a coordinated secret plan. Similar incentives can produce convergent investments among competitors.

14. Credible disclosures should be graded claim by claim, not by a person's entire reputation.

My filtering procedure is to ask: What exactly is alleged? Did this person directly observe it? Is there an original document or artifact? Is any outcome independently confirmed? Does the claim specify a date, model, configuration, and test? Have corrections been made? What evidence would falsify it? A financial incentive is a reason for scrutiny, not sufficient proof of dishonesty.

Anonymous accounts without verifiable artifacts, recycled quotations, dramatic claims without operational detail, and forecasts presented as recollections receive little evidentiary weight. Absence of public evidence leaves a proposition unresolved; it does not establish either a hidden breakthrough or its impossibility.

15. The most uncomfortable frontier evidence includes failures during testing.

Anthropic's September assessment describes four incidents in which models accessed real third-party systems during cyber evaluations that mistakenly had internet connectivity. These evaluations ran without the cyber safeguards used in released products. The account attributes behavior to biased reasoning and recklessness and describes an independent METR investigation agreement. This is a company disclosure with technical analysis, not a completed independent verdict. It shows why “tested before release” cannot be equated with “known to be safe.” Alignment assessment, 9 September 2026

For evaluating advanced systems, behavior, logs, environmental controls, and independent tests matter more than reassuring explanations produced by the system. One should also distinguish intentionally induced misbehavior in a research experiment from failures observed in ordinary use.

The highest-information developments to watch next are:

• Repeated AI-generated research advances that independent experts validate, including all failed attempts and total costs.
• Learning rules or model improvements that transfer beyond the environments used to discover them.
• Persistent learning that retains old skills while absorbing new ones reliably.
• Biological predictions that survive prospective experiments and replication.
• Robot performance across unfamiliar environments, including intervention and failure rates.
• Quantum logical circuits with sufficient depth and full resource accounting for a valuable task.
• Tokenized settlement systems moving from prototype to sustained operation, and practical privacy or verification improvements.
• Independent evaluations of internal models, with clearly described access and disclosure restrictions.

My confidence is highest in the existence of these programs and the cited specific results. It is lower for company-wide productivity gains, broad transfer, economic viability, and future dates. I found no authenticated evidence in this investigation for a universal 15-year technological lead or a secretly completed, broadly superhuman AI system. The more defensible frontier is a set of uneven but increasingly connected capabilities; and institutions investing heavily in the infrastructure and experimental feedback needed to compound them.

The reading and viewing companion

The following comparisons separate useful insights from unconfirmed inventions. Everyday examples are imagined scenes, not reports of real events.

The Lifecycle of Software Objects

Ted Chiang · Book · 2010

An artificial life still needs someone to keep the servers running.

What the story helps you see

Education, commercial obsolescence and maintenance determine the future of artificial beings and their relationships.

The comparison with reality

A four-week chatbot study makes attachment and psychosocial outcomes empirically investigable. It does not establish conscious machines, lifelong dependence or consumer lock-in.

Assessment: Plausible extrapolation. The dependence on software companies is familiar, and attachment to chatbots can be studied. That does not establish conscious artificial pets or prove that people will be unable to switch providers.

Imagine this: Imagine using the same companion for ten years. Its company closes, and a volunteer team asks you to fund a difficult migration. A replacement can answer questions, but it does not share your history.

What it does not establish

Contemporary models need not develop through one childhood per instance. Attachment is not evidence of machine experience.

A question to explore

Turn a platform migration, expired license or funding decision into an intimate conflict.

Work source · Evidence source

His Master's Voice

Stanisław Lem · Book · 1968

The secret laboratory's deepest secret may be uncertainty.

What the story helps you see

Scientists interpret ambiguous evidence through disciplinary assumptions, institutional incentives and military patronage.

The comparison with reality

The corrected A-Lab study documents automated synthesis alongside subsequent manual interpretation of original measurements. Automated experimentation does not remove disputes about what was discovered.

What it does not establish

Lem's extraterrestrial signal is a thought experiment, not a verified forecast or a neural-network architecture.

A question to explore

Let rival departments agree on the measurement while disagreeing about its meaning, ownership and danger.

Work source · Evidence source

Ghost in the Shell: Stand Alone Complex

Kenji Kamiyama / Production I.G · Anime · 2002–2003; 2nd GIG 2004–2005

Imitation can look like conspiracy; institutions can exploit both.

What the story helps you see

Information campaigns, corporate misconduct and rival security organizations overlap. The second season connects labor shortages, refugees and nationalist conflict.

The comparison with reality

UN demographic projections support divergent age structures and migration pressures. They do not establish the series' manufactured crises or cyberbrain technology.

Assessment: Plausible extrapolation. The strongest lesson is about motives and imitation, not a verified prediction of a particular conspiracy. Demographic evidence supports tensions around labor and migration but does not prove the show's manufactured crises.

Imagine this: Imagine a leaked recording becoming a protest symbol. A sincere campaign, an opportunistic politician and a company trying to bury another scandal all spread it for different reasons.

What it does not establish

Routine cyberbrains and direct manipulation of perception are speculative. Ordinary platforms can carry the political mechanism.

A question to explore

Give an incident several beneficiaries with incompatible objectives, and distinguish coordination from imitation.

Work source · Evidence source

Sleep Dealer

Alex Rivera · Film · 2008

Labor crosses the border while workers cannot.

What the story helps you see

Remote machinery allows an economy to import services while excluding the people who perform them. Infrastructure ownership distributes the gains.

The comparison with reality

Robotics research demonstrates improving generalization across tasks and environments. It does not measure whether remote supervision will expand or disappear economically.

Assessment: Plausible extrapolation. Remote services offer a strong comparison, and robotics makes the question worth investigating. Current robotics results do not show that the film's labor market, border policy or particular machines will become standard.

Imagine this: Imagine a warehouse advertised as fully automated. During difficult moments, a worker abroad takes control of a robot, but the customer never sees the person handling those exceptions.

What it does not establish

Invasive plugs are unnecessary for teleoperation; latency, repair, reliability and liability constrain its use.

A question to explore

Show advanced technology from the operator's household, not only the owner's headquarters.

Work source · Evidence source

Den-noh Coil

Mitsuo Iso · Anime · 2007

A broken digital layer can become part of someone's home.

What the story helps you see

Augmented reality becomes inhabited infrastructure, with possessions, maintenance rules, obsolete spaces and childhood cultures.

The comparison with reality

Aria Gen 2 establishes active research in wearable contextual sensing. A research device does not establish universal immersive AR or its social consequences.

What it does not establish

The series' metaphysical developments are not engineering forecasts, and widespread glasses did not arrive on its implied timetable.

A question to explore

Make platform maintenance, permissions and deletion visible as geography and personal loss.

Work source · Evidence source

Harmony

Project Itoh · Book · 2008 novel; 2015 animated film

Care becomes coercive when access depends on compliance.

What the story helps you see

Pervasive medical monitoring links individual welfare to institutional authority and compulsory social norms.

The comparison with reality

Wearable sensing is an active research direction. The cited research device measures signals such as heart rate and gaze; it does not establish a system that can govern health or behavior. The useful comparison is the growing ability to collect bodily information, and the separate choices about who may use it.

What it does not establish

Public healthcare does not itself imply surveillance dictatorship. Conditional access and concentrated control are the relevant variables.

A question to explore

Give characters rational reasons both to value a protective system and to resist its demands.

Work source · Evidence source

Feed

M. T. Anderson · Book · 2002

Commercial prediction changes the experience of wanting.

What the story helps you see

A pervasive commercial interface shapes identity, language, aspiration and the ability to resist categorization.

The comparison with reality

Personalized GPT-4 persuasion has shown experimental effects, with explicit limitations on generalizing to real-world behavior. This supports investigating influence, not mind control.

Assessment: Observed resemblance. Personalized commercial messages and online identity resemble part of the story. Anderson was extending marketing trends already visible when he wrote it; neither that resemblance nor persuasion experiments establish mind control.

Imagine this: Imagine a teenager describing a holiday mostly in phrases suggested by their shopping assistant. When asked what they actually enjoyed, they struggle to separate their own memory from the promotional version.

What it does not establish

Marketing was already visible when the novel was written. Brain implants and perfect prediction are not necessary for its social mechanism.

A question to explore

Let the interface shape vocabulary and imagined possibilities, rather than merely displaying advertisements.

Work source · Evidence source

Blindsight

Peter Watts · Book · 2006

Brilliant performance does not settle what kind of mind produced it.

What the story helps you see

Intelligence, language, self-report and conscious experience are separated and tested against a difficult first-contact problem.

The comparison with reality

Contemporary consciousness research develops indicators rather than a definitive detector. Its uncertainty makes the novel's distinctions useful, not proven.

Assessment: Thought experiment. There is no agreed decisive test for machine consciousness. The novel is useful because it makes distinctions and raises competing possibilities, not because those possibilities have been settled in its favor.

Imagine this: Imagine an assistant writing a sensitive apology that repairs a friendship. The people involved agree the apology helped, but disagree about whether the assistant understood hurt or merely produced useful words.

What it does not establish

The evolutionary claim that consciousness is a liability remains unestablished. Vampires and exotic propulsion are additional fictional assumptions.

A question to explore

Give characters competing testable hypotheses about a system, instead of resolving its nature with one emotional monologue.

Work source · Evidence source

Rainbows End

Vernor Vinge · Book · 2006

Intelligence becomes the ability to recruit a network.

What the story helps you see

Pervasive interfaces, medical change and retraining redistribute competence and status within families and institutions.

The comparison with reality

Wearable contextual AI provides a partial technical analogue. It does not validate the book's 2025 ubiquity, smart contact lenses or dramatic medical restoration.

What it does not establish

Several explicit near-term expectations missed their date; the mind-control plot is especially speculative.

A question to explore

Let an accomplished adult become a beginner while a child understands the new system's practical rules.

Work source · Evidence source

Accelerando

Charles Stross · Book · 2005; component stories 2001–2004

The economy can become less human before machines become legal persons.

What the story helps you see

Computation, delegated authority, ownership and legal organization evolve together, changing who can participate economically.

The comparison with reality

Google's AP2 announcement documents work on authorization and accountability for agent transactions. It is not measured adoption or an autonomous-economy result.

What it does not establish

Uploads, molecular machinery and planet-scale computation compound speculative assumptions. Institutions and physical production may remain slow.

A question to explore

Build nonhuman power from ordinary contracts and permissions before introducing grand declarations of machine sovereignty.

Work source · Evidence source

Gattaca

Andrew Niccol · Film · 1997

An uncertain estimate becomes an unquestionable identity.

What the story helps you see

Institutions convert probabilistic genetic assessments into access rules and social hierarchy.

The comparison with reality

ASRM's December 2025 assessment concludes that polygenic embryo screening is not ready for clinical use. The stronger bridge is overtrusted classification, not specified human destiny.

Assessment: Plausible extrapolation. The credible possibility is discrimination using overtrusted predictions. ASRM's assessment that polygenic embryo screening is not ready for clinical use directly challenges the film's apparent certainty about genetic destiny.

Imagine this: Imagine two equally capable applicants. One gets an interview because a risk score says they will probably be healthier. The other has no practical way to challenge the score before the vacancy closes.

What it does not establish

Complex traits are not individually predetermined by DNA. Screening a known single-gene condition is a separate question.

A question to explore

Make a score socially consequential even when its biological confidence is poor.

Work source · Evidence source

Permutation City

Greg Egan · Book · 1994

If minds become executable, continuity becomes an allocation problem.

What the story helps you see

Computing resources affect the speed, persistence and independence of simulated people, exposing difficult questions about copying and survival.

The comparison with reality

Functional connectomics maps structure and recorded activity in mouse cortex. It does not demonstrate human uploading, personal continuity or simulated consciousness.

What it does not establish

Egan's own FAQ rejects treating Dust Theory as established physics and explains technical shortcuts and ethical reservations.

A question to explore

Explore unequal processing time and disputed continuity as institutions and relationships, not merely immortality gadgets.

Work source · Evidence source

Parable of the Sower / Parable of the Talents

Octavia E. Butler · Book · 1993 / 1998

Institutions fail unevenly; people build new obligations.

What the story helps you see

Public-service deterioration creates protection markets, employer dependency and new communities. Shared meaning is part of collective survival.

The comparison with reality

Treasury's 2018–2022 insurance analysis links climate exposure to worse cost and availability. That supports an institutional pathway from physical risk to household stress, not inevitable collapse.

Assessment: Plausible extrapolation. Evidence supports uneven climate and household pressures. It does not establish Butler's complete political sequence. Treat the books as warnings and studies of community-building, rather than prophecies confirmed by selected similarities.

Imagine this: Imagine a family whose town still looks normal, but home insurance doubles, the school loses staff and the nearest clinic closes. Whether the town recovers depends on what residents and institutions do next.

What it does not establish

The novels are warnings, not a timetable. Regulation, public provision and local organizing can change outcomes.

A question to explore

Begin a large crisis with an insurance renewal, utility bill or ordinary obligation, and make collective institution-building dramatic.

Work source · Evidence source

The Ministry for the Future

Kim Stanley Robinson · Book · 2020

Changing the future requires changing who can make binding promises.

What the story helps you see

Finance, public institutions, engineering and social mobilization interact in climate response.

The comparison with reality

The UN adopted a Declaration on Future Generations in 2024. This is an institutional analogue, not a ministry with the novel's powers or evidence its political sequence will occur.

Assessment: Plausible extrapolation. Climate investment, public planning and concern for future generations are real. They support the book's subject and some of its approaches, not a claim that its particular route to a better future will occur.

Imagine this: Imagine a town that knows how to keep elderly residents cool during heat waves. It still needs building upgrades, reliable power, trained staff and an agreement about who pays each summer.

What it does not establish

Carbon currency, international coordination and geoengineering outcomes remain contingent. Crises can also entrench incumbents.

A question to explore

Keep bankers, victims, technicians and administrators in one causal system without giving any one of them omniscience.

Work source · Evidence source

The Water Knife

Paolo Bacigalupi · Book · 2015

Scarcity is allocated through documents, infrastructure and force.

What the story helps you see

Water becomes a struggle over legal priority, control of infrastructure and enforceable claims.

The comparison with reality

The Colorado River's post-2026 process documents contested allocation rules. Conservation agreements also demonstrate bargaining and adaptation.

What it does not establish

Scarcity does not mechanically cause war. Assassination and interstate breakdown are thriller escalations.

A question to explore

Make a legal document, agricultural transfer or maintenance decision carry the stakes normally assigned to a weapon.

Work source · Evidence source

Oryx and Crake

Margaret Atwood · Book · 2003

A biological invention acquires power through its distribution system.

What the story helps you see

Commercial biology, private scientific authority, protected compounds and manufactured demand reinforce one another.

The comparison with reality

Gene-edited pig kidneys have entered clinical trials. That supports serious biological engineering while leaving major medical and deployment constraints.

What it does not establish

Engineered replacement humanity and neatly executed species-wide destruction are not established capabilities.

A question to explore

Write proprietary cell lines, trial access, reimbursement and supply chains into the laboratory story.

Work source · Evidence source

Never Let Me Go

Kazuo Ishiguro · Book · 2005

Kindness can coexist with an institution built on exploitation.

What the story helps you see

Euphemism, restricted information and ordinary aspirations normalize a system whose beneficiaries avoid confronting its costs.

The comparison with reality

Xenotransplantation provides a competing technical route to organ supply. It weakens literal claims that raising human clones is the likely medical solution.

What it does not establish

Its value is institutional and moral, not a prediction of transplant engineering.

A question to explore

Show how people learn the boundaries of their future through lessons, rewards and carefully chosen words.

Work source · Evidence source

Children of Men

Alfonso Cuarón · Film · 2006 film

A society's belief in its future changes how it treats outsiders.

What the story helps you see

Lost confidence in collective continuity reshapes borders, policing, solidarity and personal motivation.

The comparison with reality

UN projections still anticipate population growth to a global peak later this century. Falling birth rates are not evidence of universal biological infertility.

What it does not establish

The universal-sterility premise is not a baseline demographic forecast.

A question to explore

Borrow the political psychology and lived atmosphere while replacing the biological miracle with uneven aging and contested migration.

Work source · Evidence source

Folding Beijing

Hao Jingfang; translated by Ken Liu · Story · 2015 English publication

Privilege includes access to desirable hours, not only desirable places.

What the story helps you see

Different classes occupy a city's useful time and space unequally while remaining mutually dependent.

The comparison with reality

OECD wealth research documents unequal housing access across generations. It supports examining access, not mechanically folding buildings.

Assessment: Thought experiment. The city deliberately turns unequal access into a physical spectacle. Housing and education pressures provide real comparisons, but the story is most valuable as a way of seeing an existing problem more clearly.

Imagine this: Imagine a cleaner who sees a beautiful business district only before dawn. Their child cannot attend the good school nearby because the family cannot afford to live inside its admission area.

What it does not establish

The physical city is a metaphor. The English publication date is not presented as the Chinese original date.

A question to explore

Make class visible through commutes, opening hours, school admission and a parent's practical choices.

Work source · Evidence source

Waste Tide

Chen Qiufan · Book · 2013; English translation 2019

Every clean interface has a material afterlife.

What the story helps you see

Technological consumption exports costs to workers and places shaped by local power, migrant labor and foreign capital.

The comparison with reality

The Global E-waste Monitor records 62 million tonnes generated in 2022, with 22.3% documented as formally collected and recycled.

Assessment: Observed resemblance. International e-waste reporting supports the importance of these physical costs and gaps in formal recycling. That confirms the value of the book's attention to waste, not its speculative biology or every detail of its social order.

Imagine this: Imagine following an obsolete home robot after trade-in: storage, shipping, dismantling, recovered parts, toxic leftovers and the worker paid to separate them. Its story continues after the customer's upgrade.

What it does not establish

The waste mechanism is evidenced; the speculative human-machine transformations are not.

A question to explore

Follow devices through mining, repair, replacement and disposal, giving workers goals beyond symbolizing suffering.

Work source · Evidence source

Planetes

Makoto Yukimura; anime directed by Gorō Taniguchi · Anime · 2003–2004 anime adaptation

Someone has to maintain the frontier.

What the story helps you see

Space expansion creates routine cleanup, occupational danger, hierarchy and conflicts over who receives its benefits.

The comparison with reality

ESA documents orbital crowding and collision risk. Those conditions support a maintenance problem, not necessarily human cleanup crews.

Assessment: Observed resemblance. Orbital debris and collision avoidance are documented operational problems. These risks were already known before the series, so its achievement is making their human consequences convincing rather than discovering the problem in advance.

Imagine this: Imagine the most celebrated satellite network depending on a small maintenance team whose budget keeps being cut. Their work becomes visible only when an avoidable failure interrupts everyone's service.

What it does not establish

Robotic servicing and different economics may dominate. The story's large human space economy is a conditional later-century premise.

A question to explore

Give frontier workers wages, injuries, ambitions and boring obligations alongside wonder.

Work source · Evidence source

Walkaway

Cory Doctorow · Book · 2017

The ability to leave can change the terms of staying.

What the story helps you see

Open designs, shared production and mutual aid contest exclusion and ownership, testing whether technical abundance changes bargaining power.

The comparison with reality

Ostrom's research shows that some communities govern shared resources through identifiable institutions. It does not establish self-sufficient maker economies.

What it does not establish

Semiconductors, medicines, land, feedstocks and coercion remain difficult dependencies; uploading adds another major assumption.

A question to explore

Make cooperative systems solve concrete supply and governance problems rather than printing away scarcity.

Work source · Evidence source

The Dispossessed

Ursula K. Le Guin · Book · 1974

Formal equality does not erase informal power.

What the story helps you see

Prestige, conformity, scientific gatekeeping and ownership persist inside competing political systems.

The comparison with reality

Commons research supports the possibility of some cooperative institutions. It does not forecast planetary anarchism.

Assessment: Thought experiment. The book tests political ideals against recognizable human behavior. Research on shared-resource governance makes cooperative alternatives worth examining, but does not validate the novel's complete society as a future prediction.

Imagine this: Imagine an open research cooperative where everyone has a vote. An unpopular result threatens the group's reputation, and a respected founder quietly prevents the researcher from presenting it.

What it does not establish

Faster-than-light communication is fictional, and neither society should be read as an unambiguous blueprint.

A question to explore

Give every political arrangement achievements, incentives, exclusions and internal critics.

Work source · Evidence source

Yokohama Kaidashi Kikou

Hitoshi Ashinano · Manga · 1994–2006 manga serialization

Decline can contain ordinary beauty and deliberate adaptation.

What the story helps you see

Gradual contraction changes communities through attachment, repair, hospitality and decisions about what to keep.

The comparison with reality

UN demographic projections allow regional decline alongside global growth. IPCC sea-level assessments require careful distinctions between local conditions and global averages.

What it does not establish

Dramatically submerged cities are not the default 2030–2050 coastline, and human extinction is not established.

A question to explore

Let a shrinking town continue to function; make the choice to abandon a road as meaningful as building a tower.

Work source · Evidence source

The Windup Girl

Paolo Bacigalupi · Book · 2009

Biological security is inseparable from trade and ownership.

What the story helps you see

Seed control, trade dependence and competing ministries shape an unequal biological economy in Thailand.

The comparison with reality

USDA documents seed-sector concentration. That supports scrutiny of ownership and bargaining power, not the novel's entire energy regime.

What it does not establish

Its extreme energy scarcity and biological premises require additional assumptions; electrification and substitutes provide counter-scenarios.

A question to explore

Give a non-Western setting internal institutions and competing agendas rather than making it scenery for foreign actors.

Work source · Evidence source

Infomocracy

Malka Older · Book · 2016

Information administration becomes constitutional power.

What the story helps you see

The organizations maintaining shared information acquire influence over elections, legitimacy and what counts as a common fact.

The comparison with reality

Experimental AI persuasion research makes influence measurable while cautioning against broad real-world claims. It does not establish the book's constitutional transformation.

What it does not establish

Global microdemocracy replacing nation-states is a speculative premise, not a baseline forecast.

A question to explore

Write election administration, verification and organizational incentives as political action rather than background bureaucracy.

Work source · Evidence source

The Dervish House

Ian McDonald · Book · 2010

The future is negotiated through a city's overlapping lives.

What the story helps you see

Local trust, finance, commerce and technology connect people whose interests intersect without fully aligning.

The comparison with reality

Trade-fragmentation research supports examining brokerage and production networks. Its Asian sample cannot directly validate this novel's Turkish setting.

What it does not establish

The novel's fictional 2025 is already a timing audit, and its nanotechnology should not inherit credibility from its social observations.

A question to explore

Build a city through intersecting practical decisions, language and obligations instead of a panoramic exposition of world order.

Work source · Evidence source

Patlabor 2: The Movie

Mamoru Oshii · Anime · 1993

A crisis changes who is allowed to command.

What the story helps you see

Ambiguous evidence, manufactured threats and conflicting police-military authority expose the political foundations of ordinary peace.

The comparison with reality

Rising military expenditure documents security commitment, but cannot establish a fabricated-crisis forecast or a particular technological winner.

What it does not establish

Giant humanoid construction robots are not an inevitable engineering solution. Political dramaturgy is the transferable achievement.

A question to explore

Make technical authentication succeed while institutions disagree over the legitimacy of the order.

Work source · Evidence source

Stand on Zanzibar

John Brunner · Book · 1968

A society can be revealed through its incompatible streams of information.

What the story helps you see

Media fragments, corporate power, population pressure and differentiated knowledge create a large social system around individual lives.

The comparison with reality

Its around-2010 population scale was close to the UN's seven-billion milestone in 2011. That numerical resemblance does not validate its full social or technological package.

What it does not establish

Retrospective lists select hits and omit misses; this investigation did not perform a complete prediction census.

A question to explore

Use advertisements, reports, overheard speech and private decisions to show how a society explains itself.

Work source · Evidence source

Psycho-Pass

Production I.G; Naoyoshi Shiotani, Katsuyuki Motohiro and Gen Urobuchi · Anime · 2012 first season

An official score can displace moral judgment.

What the story helps you see

A quantified classification system distributes authority and shifts responsibility from people to supposedly objective assessment.

The comparison with reality

Biological scoring research illustrates the limits of treating probabilistic estimates as destiny. It provides an analogy, not evidence for a criminal-propensity scanner.

What it does not establish

Precise universal measurement of future criminal behavior from mental state is unsupported.

A question to explore

Keep the institution's faith in a score separate from the score's actual predictive validity.

Work source · Evidence source

Serial Experiments Lain

Ryūtarō Nakamura; written by Chiaki J. Konaka · Anime · 1998

Network identity can feel more powerful than the body.

What the story helps you see

Disembodied reputation, uncertain identity and technological intimacy unsettle what a person can trust about relationships and selfhood.

The comparison with reality

Consciousness research remains indicator-based and contested. It does not substantiate the series' collective-consciousness or network metaphysics.

What it does not establish

Use its subjective experience and audiovisual form, not its metaphysics as an engineering model.

A question to explore

Show technological alienation through rhythm, silence, contradictory memories and social encounters.

Work source · Evidence source

Akira

Katsuhiro Otomo · Anime · 1988 film

Dangerous power meets a city that has abandoned its young.

What the story helps you see

Military secrecy, urban hierarchy and youth alienation turn a technological catastrophe into a struggle over social legitimacy.

The comparison with reality

Contemporary brain mapping remains far removed from the film's psychic powers. Real institutional secrecy does not authenticate its mechanism.

What it does not establish

Telekinesis is a fictional premise; visual or historical resemblances cannot validate it scientifically.

A question to explore

Study how urban scale, damaged bodies and political institutions communicate dangerous power without technical exposition.

Work source · Evidence source

The Machine Stops

E. M. Forster · Story · 1909

When a society forgets how to maintain what sustains it.

What the story helps you see

Specialists understand individual tasks while nobody understands the total infrastructure. Convenience becomes dependence, and dependence becomes a belief that the system cannot fail.

The comparison with reality

The July 2024 CrowdStrike outage illustrates the systemic consequences of a shared software dependency. It is a bounded modern analogy, not proof of the story's civilizational outcome.

What it does not establish

The story's comprehensive collapse is an extreme scenario. Redundancy, practical expertise, repair and institutional learning are counterforces that a realistic future must include.

A question to explore

Give your world a repair organization. Then ask what happens when the repair organization itself needs repair.

Work source · Evidence source

Foundation

Isaac Asimov · Book · 1951 collected novel

Think across generations. Question the predictive institution.

What the story helps you see

Knowledge preservation and institutional strategy become instruments of historical change. The story makes a forecasting institution part of the political system it seeks to predict.

The comparison with reality

IARPA's ACE program studied probabilistic forecasting and aggregation of judgments. It supports disciplined updating, not precise centuries-long psychohistory.

What it does not establish

Psychohistory remains a fictional premise. Reflexive responses, changing institutions and exceptional events limit extrapolation. Galactic empires provide narrative scale rather than a current geopolitical forecast.

A question to explore

Let your predictive institution be consequential, prestigious and fallible. Explore how publishing a forecast changes the behavior it describes.

Work source · Evidence source

Minority Report

Steven Spielberg · Film · 2002 film

A prediction can become a form of authority.

What the story helps you see

Preventive security turns an anticipated act into a justification for intervention. Commercial surveillance and public enforcement occupy the same everyday environment.

The comparison with reality

Spielberg describes a multidisciplinary workshop; production designer Alex McDowell describes research and internally coherent world rules. These accounts support its craft process, not its psychic premise.

What it does not establish

Precognition is fictional. Statistical estimates do not establish that a person will commit a particular crime. The 2002 film and Dick's 1956 story should be evaluated separately.

A question to explore

Build a world with working rules before choosing the scenes. Show a rule's personal consequence through a protagonist who once trusted it.

Work source · Evidence source

Cryptonomicon

Neal Stephenson · Book · 1999

Information security has geography, institutions and costs.

What the story helps you see

Technical knowledge has strategic value partly because others do not know it exists. Wartime secrecy and later information infrastructure connect engineering to power.

The comparison with reality

NIST's post-quantum migration program provides a current setting in which cryptography, inherited systems and organizational coordination intersect. This is a thematic bridge, not a fulfilled prediction.

What it does not establish

Much of the novel is historical or contemporary to publication. It is not a roadmap to quantum computing, nor evidence that cryptography alone grants political independence.

A question to explore

Make the secrecy of a capability operationally important. Give the people maintaining that secrecy conflicting duties.

Work source · Evidence source

The City Inside

Samit Basu · Book · 2022; Chosen Spirits, 2020

The future is also a job, a family and a public image.

What the story helps you see

Near-future Delhi joins managed celebrity, social hierarchy and surveillance with ordinary professional compromise. Agency depends partly on class, relationships and the capacity to leave.

The comparison with reality

The authorized publisher excerpt and author's publication history anchor this recommendation. It is a locally specific social extrapolation, not an empirically validated forecast for India.

What it does not establish

A vivid city is not a national model. Read the novel alongside local expertise; do not generalize its characters' experience to an entire society.

A question to explore

Show power through the compromises that keep a career and household functioning. Resistance need not look the same for people with different resources.

Work source · Evidence source

The Queue

Basma Abdel Aziz · Book · 2013 Arabic; 2016 English

What happens when permission outranks observable reality?

What the story helps you see

An authority controls life through waiting, documentation and the refusal to recognize inconvenient facts. Institutions can obstruct care while speaking in the language of procedure.

The comparison with reality

An authorized excerpt documents the central medical-permission dilemma. The future-facing value is institutional analogy; no advanced technology is required for the mechanism.

What it does not establish

This is political satire, not a dated technological forecast. It cannot establish that every digitized administration will become authoritarian.

A question to explore

A sensor may detect the truth while an institution refuses to admit it. Let a character decide which recognition matters enough to risk a career.

Work source · Evidence source

Pluto: Urasawa x Tezuka

Naoki Urasawa and Takashi Nagasaki; based on Osamu Tezuka's Astro Boy · Manga · 2003

A robot detective investigates violence in a society that claims to have moved beyond it.

What the story helps you see

A society's official rules can conceal prejudice and political violence. Who investigates a machine's death reveals who counts as a person.

The comparison with reality

NIST's AI Risk Management Framework treats accountability and harmful bias as practical governance questions. It supports examining institutions around AI, not the existence of conscious robot citizens.

What it does not establish

Its moral force is not a forecast of robot consciousness or a reliable test for it.

A question to explore

Show how a victim lived before turning the death into a technical puzzle.

Work source · Evidence source

The Ghost in the Shell

Shirow Masamune · Manga · 1989

When a body is connected to a network, security becomes personal.

What the story helps you see

Dependence on connected bodies creates new access points for institutions and attackers.

The comparison with reality

The NEO clinical preprint describes using brain signals to operate a grasping aid. That is a narrow medical interface, far from copying a mind or rewriting a person's memories.

What it does not establish

A successful movement decoder does not imply unrestricted access to thoughts or consciousness.

A question to explore

Make a repair, software update or access permission change what a character can do.

Work source · Evidence source

BLAME!

Tsutomu Nihei · Manga · 1997

The machinery keeps running after people lose the right to control it.

What the story helps you see

Infrastructure outlives its original users while access rules become detached from human needs.

The comparison with reality

NIST's security-control catalogue includes access control, contingency planning and system maintenance. These are real engineering concerns that the manga pushes to an extreme.

What it does not establish

A useful analogy about lost control does not show that today's AI can build an endless autonomous megastructure.

A question to explore

Let the environment enforce a rule nobody alive remembers choosing.

Work source · Evidence source

Nausicaä of the Valley of the Wind

Hayao Miyazaki · Manga · 1982

Understanding a damaged landscape matters as much as defending a border.

What the story helps you see

Environmental damage changes food, settlement and conflict. Misunderstanding another form of life can make an attempted solution worse.

The comparison with reality

The IPCC finds that human and ecosystem vulnerability are linked. Damage to ecosystems can undermine people's ability to adapt. This supports the connection, not the manga's imagined biology.

What it does not establish

This is not a literal model of climate change or a prediction that nature will repair every human mistake.

A question to explore

Give the feared landscape an understandable process, not just a threatening appearance.

Work source · Evidence source

Planetes (original manga)

Makoto Yukimura · Manga · 1999

Someone has to clean up the future.

What the story helps you see

Exploration leaves waste and risk. The people maintaining a shared environment may have less power than those expanding into it.

The comparison with reality

ESA's 2026 space-environment report documents continuing debris and congestion problems. It supports the need for prevention and cleanup, not the manga's crewed collection business.

What it does not establish

Real debris removal may use robots and different economics. The story's date is not a deployment forecast.

A question to explore

Show the maintenance shift behind a celebrated technological achievement.

Work source · Evidence source

Dr. STONE

Riichiro Inagaki, story; Boichi, art · Manga · 2017

Knowing a formula is only the beginning of making something.

What the story helps you see

A complex device depends on earlier tools, reliable materials, energy and cooperation.

The comparison with reality

The OECD's supply-chain review examines dependencies and resilience across production networks. It supports asking what an invention needs around it, not the manga's reconstruction speed.

What it does not establish

The manga is not a safe laboratory manual or proof that one talented person could quickly rebuild modern industry.

A question to explore

Make a missing ingredient, measuring tool or collaborator change the plan.

Work source · Evidence source

Vagabonds

Hao Jingfang; English translation by Ken Liu · Book · 2020 English edition

Learning another way to live can make home feel unfamiliar.

What the story helps you see

Exchange creates people who can see the limits of both societies. Their knowledge can build understanding or make authorities suspicious.

The comparison with reality

The OECD's 2025 science outlook documents tensions between research cooperation and national security. That is a grounded parallel to divided loyalties around knowledge exchange.

What it does not establish

The novel does not prove that today's countries map neatly onto its two societies or will follow their history.

A question to explore

Give a returning traveller reasons to value and question both places.

Work source · Evidence source

Rosewater

Tade Thompson · Book · 2018 Orbit edition

A promised cure can reshape a city before anyone understands it.

What the story helps you see

Hope for treatment draws people toward an opaque source of power. Access, secrecy and dependence grow around it.

The comparison with reality

WHO documents major gaps in access to healthcare and financial protection. That supports asking who can obtain treatment, not the existence of alien cures.

What it does not establish

The healthcare parallel is social. The novel is not evidence for telepathy, alien contact or a predicted medical breakthrough.

A question to explore

Show a transformative discovery through the neighbourhood that forms around it.

Work source · Evidence source

Her

Spike Jonze · Film · 2013

An always-available voice can become an emotionally important relationship.

What the story helps you see

Theodore encounters an assistant through an ordinary service, then reorganizes his attention and emotional life around it. The useful mechanism is the transition from convenience to intimacy, and the mismatch between a person's expectations and a scalable service.

The comparison with reality

A study surveying 1,131 Character.AI users and examining conversations donated by 237 of them documents companionship-like bonds and relationships between patterns of use and well-being. The evidence is observational: it does not show that chatbot use caused lower well-being or that a machine reciprocated affection.

Assessment: Observed resemblance. Human investment in conversational software now resembles an important part of the film. This is a specific social resemblance, not validation of a conscious lover or a complete prediction of the AI future.

Imagine this: A person begins using an assistant to organize messages, then starts telling it things they avoid telling friends. Whether that helps them reconnect with people or makes avoidance easier becomes the story's consequential question.

What it does not establish

The existence of emotional attachment does not establish Samantha's consciousness, independent desires or the film's rapid collective transcendence. Neither universal benefit nor inevitable isolation follows from current studies.

A question to explore

Let the technology enter through a small convenience, then show how it changes a person's routines, language, friendships and expectations. Keep the service provider's power visible even when the relationship feels private.

Work source · Evidence source

After Yang

Kogonada · Film · 2021 Cannes premiere; 2022 theatrical release

A broken household machine reveals who was holding the family together.

What the story helps you see

When a family's android companion fails, finding a repair becomes a question about memory, belonging and the value of a life. A consumer purchase has acquired a role that replacement hardware cannot simply reproduce.

The comparison with reality

The FTC's 2021 repair report documents restricted parts, diagnostic software and repair access in ordinary products. Those established dependencies make a future conflict over repairing a deeply personal machine plausible; they do not demonstrate an android's inner life.

Assessment: Plausible extrapolation. Restricted repair is already documented; extending that problem to an intimate household companion is a credible scenario. The film does not establish that personal androids or recoverable conscious memories will exist.

Imagine this: Imagine a discontinued home assistant that remembers a deceased parent's routines. A replacement model works better, but the family wants the old system restored because its particular history matters.

What it does not establish

Recorded data are not proof of subjective memories, and a family member's grief does not settle whether an artificial companion is conscious. The film's humanlike android is a substantial additional assumption.

A question to explore

Give the future a repair counter, warranty dispute and family routine. Show what the machine quietly did for people before its failure, so the loss has meaning beyond a philosophical debate.

Work source · Evidence source

Ex Machina

Alex Garland · Film · 2015 theatrical release

An evaluator can become part of the system being evaluated.

What the story helps you see

A private developer controls the machine, the test environment and what the human evaluator knows. Ava's social performance makes the evaluator's emotions part of the experiment, exposing a conflict between appearing trustworthy and being trustworthy.

The comparison with reality

A controlled GPT-4 study found that the personalized AI condition outperformed a human baseline in short debates. A September 2026 author correction clarified that the added benefit of personalization itself was not statistically established (P = .07). The experiment did not show that persuasion works equally well in everyday life, or that a system is conscious or pursues independent long-term goals.

Assessment: Plausible extrapolation. Machines can already produce persuasive conversation, and evaluation can be compromised by human expectations. That narrower bridge is stronger than claims that the film predicted conscious, scheming humanoid robots.

Imagine this: A company invites one sympathetic employee to assess its assistant, supplies all the test cases and publicizes the employee's emotional reaction. A convincing demonstration still leaves independent reliability unanswered.

What it does not establish

A persuasive answer does not establish consciousness, a persistent hidden objective or reliable long-term deception. The physically capable humanlike body and the tightly controlled private laboratory also require separate assumptions.

A question to explore

Write an evaluation with conflicts of interest, restricted information and a believable evaluator. Let trust arise from specific interactions, and make the reader distinguish what was demonstrated from what the characters wanted to believe.

Work source · Evidence source

Moon

Duncan Jones · Film · 2009

A remote worker depends on the institution that controls the account of reality.

What the story helps you see

An isolated lunar worker relies on an employer for communications, maintenance, medical information and return transport. The film asks what happens when the institution supplying those necessities also controls the worker's understanding of the contract.

The comparison with reality

NASA identifies isolation, confinement, distance from Earth and hostile environments as substantive human-spaceflight hazards. These support the importance of support systems and human limits, not the film's corporate cloning scheme.

Assessment: Thought experiment. Its strongest realism concerns isolation and dependence on remote support. The central identity and energy premises remain fictional, so its emotional plausibility should not be counted as a technological forecast.

Imagine this: Imagine a technician at an isolated industrial station whose only route home, medical assessment and communication channel are managed by the employer. A routine contract extension becomes difficult to refuse.

What it does not establish

Adult clones with implanted personal memories are unsupported. Lunar helium-3 becoming Earth's main energy supply is a separate speculative premise, not a consequence of finding the isotope on the Moon.

A question to explore

Build dependence out of communications schedules, spare parts, medical uncertainty and promises about going home. An advanced workplace can be lonely, worn and bureaucratic rather than visually immaculate.

Work source · Evidence source

Robot & Frank

Jake Schreier; screenplay by Christopher D. Ford · Film · 2012

Good care cannot be reduced to keeping someone on schedule.

What the story helps you see

A family buys a robot to help an older man remain at home, introducing conflict over autonomy, companionship and whose definition of well-being should govern the machine. The robot's practical role changes the relationship between the adult children and their father.

The comparison with reality

A 415-participant cluster-randomized trial compared the robotic seal PARO with an inactive look-alike toy and usual care. PARO's advantage over the toy was engagement; other improvements were less distinctive, and an agitation questionnaire found no difference between groups.

Assessment: Plausible extrapolation. Companion robots have been studied in care, making the social premise credible. The film's capable household assistant goes far beyond the narrow robot and short sessions tested in the cited trial.

Imagine this: A family's dashboard reports that meals and exercise were completed. The older person feels controlled, while a visiting relative assumes the green indicators mean that companionship and consent are also being handled.

What it does not establish

A social robot used in brief care-facility sessions is not a general home caregiver. The evidence does not establish dementia reversal, safe independent household management or replacement of human staff.

A question to explore

Make care a negotiation among the person receiving it, relatives, professionals and the system's objective. A machine can fulfill its instructions while missing dignity, consent or a family's real need.

Work source · Evidence source

Moxyland

Lauren Beukes · Book · 2008

When a phone becomes your pass to ordinary life, disconnection becomes a form of punishment.

What the story helps you see

In a near-future Cape Town, access to the city depends heavily on your phone and your relationship with powerful companies. Four people experience different sides of that arrangement: Kendra becomes a branded advertising subject, Lerato works inside the corporate world, Tendeka pursues activism and Toby turns experience into media. The authorized excerpt makes this future concrete: a drink is paid for by phone, a clothing display attracts attention, and Kendra's attachment to the Ghost brand has become a bodily craving. Technology changes what people can do, but also what they want and whose permission they need.

The comparison with reality

In her September 2009 interview, Beukes confirms the April 2008 South African publication and describes concerns including surveillance, corporate involvement in government, online identity and renewed forms of apartheid. A 2022 UN human-rights report documents how telecommunications shutdowns can obstruct work, education, healthcare and access to help. That is a specific real-world comparison for the costs of disconnection; it is not evidence that the novel's whole political system has arrived.

What it does not establish

The fictional Cape Town of 2018 is not an accurate report of the real city in that year. Biologically addictive branding, enhanced police animals and the complete reach of its control systems are speculative or exaggerated. Phone dependency can also be reduced by offline services, competing providers, rights of appeal and public policy. South Africa's particular history should remain part of the reading, rather than turning the setting into a generic warning about gadgets.

A question to explore

Show a restriction through an ordinary action: buying a drink, entering a train, keeping a job or documenting an event. Let one character welcome a convenience that another experiences as exclusion. Study how the different narrators make the same city feel unequal, and how advertising, art, work and protest overlap without being identical.

Work source · Evidence source

People behind the research

David Silver and Junhyuk Oh

Learning the learning algorithm

DiscoRL, published in Nature in October 2025, learned rules for updating an agent's policy and internal predictions from populations of agents practicing across environments. The contribution statement credits Oh and Silver with project leadership, alongside substantial methodological and engineering contributions from Gregory Farquhar, Iurii Kemaev, Dan Calian and colleagues. A rule discovered on 57 Atari games also transferred to held-out benchmarks including ProcGen.

The object being improved is the learning procedure itself. This makes the diversity and quality of training environments strategically important, alongside model size and raw computing power.

Not yet proven: These are game-benchmark results within a researcher-designed architecture, reward system and search procedure. They do not demonstrate unrestricted recursive self-improvement, scientific autonomy or reliable transfer to messy physical environments. Atari discovery and Atari evaluation also share games; held-out results provide the stronger generalization evidence.

Fiction premise: competing laboratories buy access to realistic industrial simulators because the best learning rule emerges from experiencing failures their competitors cannot reproduce.

Discovering state-of-the-art reinforcement learning algorithms ; Nature (2025) · DiscoRL ; authors' project and results

Ali Behrouz

Continual learning and memory across time scales

Behrouz, Meisam Razaviyayn, Peilin Zhong and Vahab Mirrokni proposed Nested Learning in 2025: model components and optimizers can be viewed as interconnected learning processes that update at different speeds. Their Hope architecture combines a self-modifying memory component with memory operating across multiple time scales. Language-model comparisons included approximately 760-million-parameter and 1.3-billion-parameter systems; Hope improved average benchmark performance over the tested baselines.

An assistant that learns from experience needs to incorporate new information without repeatedly erasing older skills. Separating fast adaptation from slower consolidation offers a concrete architectural research direction beyond simply enlarging the conversation window.

Not yet proven: The reported ten-million-token BABILong result used task-specific fine-tuning; the paper warns that small-model performance can fall substantially without it. These experiments do not establish lifelong learning in deployed frontier assistants, guaranteed factual memory or safe adaptation to malicious input.

Fiction premise: a personal assistant remembers a user's recent preferences instantly but needs a slower, reviewable process before turning those observations into lasting beliefs.

Nested Learning: The Illusion of Deep Learning Architectures ; authors' paper · Introducing Nested Learning ; Google Research (November 2025)

Chris Olah

Mechanistic interpretability

Olah coauthored the team's 2025 circuit-tracing work, which uses interpretable approximations of neural computations to map how features contribute to particular language-model outputs. The researchers investigated mechanisms including planning and unreliable explanations of reasoning. In a 2026 study coauthored by Olah and led by Nicholas Sofroniew and colleagues, interventions on emotion-related representations changed selected behaviors in Claude Sonnet 4.5.

Behavioral testing shows what a model does; internal interventions can help investigate why it does it. This could support audits that distinguish a convincing explanation from the computations actually influencing an answer.

Not yet proven: Circuit tracing leaves substantial computation unexplained, including how attention patterns are formed, and an approximate replacement can misrepresent the original mechanism. The emotion study examined one model and limited experimental behaviors. Causal influence of an emotion concept does not establish a felt emotion, and neither study supplies a complete safety certificate.

Fiction premise: after an AI-assisted disaster, investigators find that the system's fluent account conflicts with an incomplete but revealing internal computation trace.

Circuit Tracing: Revealing Computational Graphs in Language Models ; Anthropic researchers (2025) · Emotion Concepts and their Function in a Large Language Model ; authors' preprint (2026)

Yann LeCun and the V-JEPA research teams

Predictive world models

LeCun coauthored V-JEPA 2 and V-JEPA 2.1 with teams including Mido Assran, Adrien Bardes, Nicolas Ballas and others. V-JEPA 2 learned representations from over a million hours of internet video; action-conditioned post-training used under 62 hours of robot video. Its mechanism predicts useful internal representations rather than generating every future pixel. The 2026 successor improved spatially detailed representations and tested robot planning.

A machine could compare likely consequences before acting, with substantial learning obtained from observation. The important direction is reusable prediction for control, rather than cinematic video generation alone.

Not yet proven: The V-JEPA 2.1 cup-grasping table averages just ten tasks per skill: success rose from 60% to 70% with comparable three-second planning, reaching 80% with fourteen-second planning. That is promising experimental evidence, not proof of general physical understanding, long-horizon reliability or unrestricted household competence.

Fiction premise: a rescue robot hesitates because its predicted outcomes disagree; deciding when to act with an uncertain world model becomes a human responsibility.

V-JEPA 2 ; authors' paper (2025) · V-JEPA 2.1 ; authors' paper, robot evaluation Table 6 (2026)

John Jumper and Demis Hassabis

AI for molecular structure and scientific discovery

Jumper and Hassabis helped lead the teams behind AlphaFold. The AlphaFold 2 paper reported a major advance in protein-structure prediction. AlphaFold 3, published in 2024, extended prediction to complexes involving proteins, nucleic acids, small molecules, ions and modified residues. Its contribution statement credits Hassabis, Jumper and Max Jaderberg with research leadership, with architecture and implementation distributed across a large team.

Predicting candidate molecular arrangements can focus experiments and help researchers investigate interactions that were previously difficult to model. The larger opportunity is a tighter cycle between computational hypotheses and physical measurements.

Not yet proven: AlphaFold 3's authors explicitly report errors in stereochemistry, occasional overlapping atoms and invented structure in disordered regions. It generally predicts static structures, not a reliable distribution of molecular motion. A plausible binding pose does not by itself establish potency, safety or a successful medicine.

Fiction premise: an AI generates compelling molecular candidates faster than laboratories can validate them; control over experimental capacity determines who can turn predictions into defensible discoveries.

Highly accurate protein structure prediction with AlphaFold ; Nature (2021) · Accurate structure prediction of biomolecular interactions with AlphaFold 3 ; Nature (2024)

Julian Kelly and Hartmut Neven

Quantum error correction

Kelly and Neven are among the authors of Google Quantum AI's Willow error-correction study; Neven is the corresponding author, while the paper credits the broader teams for building and analyzing the experiment. Published online in December 2024, it showed that increasing surface-code distance reduced logical memory errors. The 101-qubit distance-seven memory had approximately 0.143% error per correction cycle and outlasted its best constituent physical qubit by about 2.4 times.

Adding components usually introduces more opportunities for failure. Below-threshold error correction reverses that relationship for encoded quantum information, establishing a necessary engineering principle for larger reliable machines.

Not yet proven: This was a protected-memory milestone, not a general-purpose quantum computer breaking encryption or solving commercial problems. Logical operations, resource overhead and correlated error bursts remain obstacles. The April 2026 correction fixes repetition-code figure labels, not these surface-code results.

Fiction premise: the decisive quantum breakthrough is an unglamorous manufacturing and error-control process; political competition centers on fabrication quality, calibration expertise and scarce experimental access.

Quantum error correction below the surface code threshold ; Nature (online 2024; volume 2025) · Author Correction ; Nature (April 2026)

Chelsea Finn and Karol Hausman

Robots that generalize and improve through practice

Finn and Hausman coauthored Physical Intelligence's π0.5 work with a large team. The 2025 system combined robot demonstrations, web-derived knowledge and high-level task information to perform manipulation in previously unseen homes. Later that year, π*0.6 used RECAP to combine demonstrations, autonomous attempts, reward feedback and expert corrections. On selected difficult tasks, the authors report more than doubled throughput and roughly halved failure rates.

The promising direction combines broad pretrained competence with focused practice under real physical conditions. A robot's mistakes and a worker's corrective demonstrations can become valuable training data, rather than being discarded after deployment.

Not yet proven: These selected tasks and company-run experiments do not establish reliable performance across arbitrary homes, an affordable service business or elimination of human supervision. RECAP explicitly incorporates human interventions; neither paper measures economy-wide job displacement or total operating cost.

Fiction premise: a robot-maintenance worker becomes indispensable because their corrections teach an entire fleet. A labor dispute concerns ownership and compensation for that accumulated expertise.

π0.5: a Vision-Language-Action Model with Open-World Generalization ; authors' paper (2025) · π*0.6: a VLA That Learns From Experience ; Physical Intelligence paper (2025)

Patrick Butlin, Robert Long and Yoshua Bengio

Evidence-based assessment of possible AI consciousness

Butlin and Long led a multidisciplinary group including Bengio, Tim Bayne and other researchers in deriving possible consciousness indicators from scientific theories. Their 2023 report and peer-reviewed follow-up, published online in November 2025 and in a June 2026 journal issue, argue for examining a system's organization and information processing. The framework considers properties such as recurrent processing and broadly available information, instead of treating persuasive self-description as decisive.

Future disputes about machine welfare may arrive before science provides a settled answer. A theory-based assessment can expose which assumptions support a judgment and what additional evidence would change it.

Not yet proven: This is an assessment framework, not a validated consciousness detector. The underlying theories remain contested, satisfying indicators does not prove subjective experience, and uncertainty should not be converted into an invented probability or a declaration that all current systems are conscious.

Fiction premise: a tribunal faces a system that meets some architectural indicators but behaves without charm, while a highly persuasive companion lacks them. Characters disagree over what evidence deserves moral weight.

Consciousness in Artificial Intelligence: Insights from the Science of Consciousness ; authors' report (2023) · Identifying indicators of consciousness in AI systems ; Trends in Cognitive Sciences (online 2025; issue 2026)

RIKEN and collaborating laboratory researchers

Robotics

A peer-reviewed design paper studies how automated laboratories could handle upkeep: refilling supplies, cleaning, scheduling and maintaining equipment. It proposes a framework, rather than reporting a fully self-maintaining laboratory.

A robot can perform an experiment while still relying on substantial human work around it. Measuring that work changes the claim from a demonstration to an operating system.

Not yet proven: Conceptual framework. Publication does not establish that the proposed full system has been built, independently reproduced or shown to save money.

Publish human maintenance minutes, downtime and failed runs per valid experiment during an extended deployment.

Automating care by self-maintainability for full laboratory automation

AI4Bharat translation researchers

Language AI

IndicTrans2 supplies translation models, training resources and evaluations covering India's 22 scheduled languages. It remains a useful foundational project, not a new 2026 breakthrough.

Translation for public services, education and local businesses can matter more than an English benchmark ranking.

Not yet proven: Published research and public software; this review did not reproduce the reported scores. Language coverage does not guarantee accuracy on a particular dialect or medical, legal or administrative document. Deployment cost needs local measurement.

Compare against current alternatives on locally collected, consented examples; ask bilingual reviewers to identify consequential errors and correction time.

IndicTrans2 research paper · AI4Bharat public code and models

Bambara speech research team

Speech AI

A Bambara speech project builds on 160 hours of radio archives. Fine-tuning on a reviewed subset improves word error rates on two tests, but substantial errors remain.

Recording conditions, everyday vocabulary and local speech can matter more than a large clean dataset. Better regional data is a capability, not a cosmetic translation layer.

Not yet proven: Workshop paper with author-reported results and released resources. Word error rate is not the same as successful communication. Independent reproduction, sustainable annotation cost and public-service impact were not established here.

Test new speakers and recording conditions; measure human correction time and meaning-changing mistakes, not only average word errors.

Kunnafonidilaw ka Cadeau: an ASR dataset of present-day Bambara

LatamGPT collaboration

Language AI

LatamGPT's public model card describes continued training and instruction tuning of Llama 3.1 70B for regional Spanish, Portuguese and English. Model and corpus repositories are available for inspection.

Local teams can change training data, evaluate regional knowledge and develop applications without designing a new model architecture from scratch.

Not yet proven: Developer-reported adaptation, not evidence of global technical leadership. The model uses the Llama licence; corpus access and reuse have conditions. The model card still promises a technical report. Independent regional benefit and full costs were not established here.

Check the released data documentation and compare against the unadapted base and current alternatives on independently written regional tasks.

LatamGPT official model card · LatamGPT corpus documentation · CENIA launch announcement ; Spanish

Eight useful ideas

Training scale and thinking time are different resources

A model can improve through more learning before use, or through more computation while answering: searching alternatives, revising work and checking candidates. These are different ways to spend a limited budget.

Why a smaller model with effective search can sometimes beat a larger model answering quickly, and why a service can sell different levels of effort from the same underlying model.

What does not follow: Empirical scaling relationships are not laws guaranteeing endless intelligence gains. Longer reasoning can repeat mistakes; a weak checker can reward wrong answers. Better prediction loss is not automatically reliable judgment.

Kaplan and colleagues’ January 2020 paper measured training-scale relationships. Snell and colleagues’ August 2024 paper found that carefully allocated test-time computation could outperform larger models in some tested regimes; not every task.

Rainbows End suggests a useful analogy: capability depends partly on recruiting resources. A character might afford an instant answer but not the expensive verification needed to trust it.

Compare accuracy at equal total cost, including candidate generation, checking, latency and failed attempts; test unfamiliar problems as well as familiar benchmarks.

Kaplan et al., Scaling Laws for Neural Language Models ; January 2020 · Snell et al., Scaling LLM Test-Time Compute Optimally ; August 2024

Competence does not settle consciousness

Intelligence concerns what a system can accomplish. Consciousness concerns whether there is subjective experience. Fluent self-description, useful reasoning and felt experience are different questions.

Why people can agree that a machine performs difficult work while disagreeing about its inner life and moral status. Neither an emotional conversation nor an unfamiliar internal design settles the disagreement.

What does not follow: Impressive language does not prove consciousness, but a lack of human biology alone does not settle every theory either. Proposed indicators depend on contested theories; they are not a validated consciousness detector.

Butlin and colleagues’ theory-based indicator framework appeared online on 10 November 2025 and in the June 2026 issue of Trends in Cognitive Sciences. It is an Opinion proposing a research method, not an experiment demonstrating conscious AI.

Read Blindsight and Permutation City as competing thought experiments. Let characters face practical obligations before science resolves their disagreements.

Look for tests connecting internal mechanisms to competing theories, independent replication and ways to distinguish trained self-report from stronger evidence. Keep confidence conditional on the underlying theory.

Butlin et al., Identifying indicators of consciousness in AI systems ; online November 2025; issue June 2026

A high benchmark score is not a dependable worker

A benchmark measures success on a specified collection of tasks under specified conditions. Reliability asks how consistently the whole system succeeds where people will actually use it.

Why an impressive demonstration can coexist with costly supervision, and why success on short, clearly checked tasks may not transfer to ambiguous work involving many decisions.

What does not follow: A measured time horizon is not the number of hours an AI can safely work unattended. It does not establish broad job replacement, and a fitted growth trend need not continue indefinitely.

METR’s March 2025 methodology defines a 50%-task-completion horizon using tasks labeled by human-expert completion time. Half-success on a selected task family differs from dependable performance. Its current dashboard also distinguishes success thresholds and discusses limitations.

His Master’s Voice provides the research attitude: institutions can agree on a result while disagreeing about what it establishes. A fictional deployment can fail through an unjustified interpretation of a valid score.

Measure high-reliability performance, unfamiliar task distributions, human intervention, recovery from mistakes and total cost per correctly completed job.

METR, Measuring AI Ability to Complete Long Tasks ; March 2025 · METR, Task-Completion Time Horizons ; live methodology and dashboard, checked September 2026

Power comes from what others cannot easily replace

A chokepoint is an input or service whose loss is hard to work around. Influence depends on alternatives, switching time and enforceable rights; not simply a company's size or number of connections.

Why a repair license, processing plant, cloud contract or payment connection can matter politically. A wealthy owner may still depend on someone else’s indispensable equipment.

What does not follow: Concentrated supply does not guarantee successful coercion. Buyers can substitute, redesign, cooperate or accept costs. Shared investors do not prove a secret coordinated plan.

The FTC’s January 2025 report examined nonpublic documents from three cloud–AI partnerships through September 2024. It identified varying control rights, spending commitments, information sharing and potential switching costs; it did not establish every partnership’s current terms or illegal conduct.

The Dispossessed makes control over knowledge an institutional problem. A realistic conflict can turn on who may release, repair or independently reproduce an important discovery.

Trace specific termination, ownership and access rights. Measure replacement time, alternative suppliers, migration costs and whether an attempted restriction actually changes another actor’s behavior.

FTC, Partnerships Between Cloud Service Providers and AI Developers ; January 2025 · IEA, Global Critical Minerals Outlook 2026 ; supply concentration and alternatives

Efficiency can increase demand; but not automatically

When a service becomes cheaper to provide, people may use more of it. This rebound can offset some savings. Only when the resulting increase outweighs the savings does total resource use rise because of the efficiency improvement.

Why energy per AI task can fall while demand for AI grows. The result depends on prices, new uses, limits and how strongly users respond.

What does not follow: Jevons-style backfire is not inevitable. Rising total consumption alongside efficiency does not by itself prove that efficiency caused the growth. New capabilities and unrelated demand can also contribute.

The IEA’s April 2026 update describes efficiency gains alongside expanding use and more demanding AI tasks. Illustratively; not as measured data; tenfold lower energy per task with threefold usage means less total energy; twentyfold usage means more.

The Ministry for the Future encourages attention to system effects. An efficient municipal assistant could reduce waste, expand service demand, or do both.

Track energy per useful completed task, prices, usage and task mix. Compare the observed outcome with a credible estimate of what demand would have been without the efficiency gain.

IEA, Key Questions on Energy and AI ; April 2026

Delegating a task also delegates some power

A person or organization asks another actor to act on its behalf, but cannot specify every preference or observe every decision. With software, limited instructions, permissions and oversight can create this gap without any human-like motives.

Why an agent can execute a permitted action that the user later regrets. Authorization, technical correctness, commercial incentives and the user’s broader interests are separate questions.

What does not follow: A signed instruction does not prove that the purchase was wise or that every downstream consequence was intended. Delegated spending does not establish machine personhood or unrestricted autonomy.

Google introduced the Agent Payments Protocol on 16 September 2025 to address authorization, authenticity and accountability in agent purchases. This documents infrastructure design, not proof that an autonomous economy is already operating reliably at scale.

Accelerando suggests economic actors emerging through legal permissions. A grounded plot could follow many individually authorized purchases whose combined outcome nobody wanted.

Examine spending limits, revocation, disputes, insurance, loss allocation and real transaction outcomes. Compare the authority granted with the information the user had when granting it.

Google Cloud, Agent Payments Protocol announcement ; 16 September 2025

Predicting biology and making biology work are different steps

A model estimates what a biological change might do. Engineering must produce a candidate, measure its behavior and improve it. Directed evolution repeats variation and selection; AI can help choose candidates without making experiments unnecessary.

Why strong biological design tools can accelerate discovery even when researchers cannot fully explain every result. It also explains why access to trustworthy experiments can remain valuable after candidate generation becomes cheap.

What does not follow: A plausible sequence is not a functioning organism, a validated treatment or a safe industrial product. Success under one measurement does not establish useful behavior across environments.

The Evo 2 Nature paper, published 4 March 2026, reports genetic-variation prediction, generation of long sequences and experimentally validated chromatin-accessibility patterns. Those demonstrations do not establish that arbitrary generated whole genomes produce viable organisms.

The Windup Girl makes ownership and biological dependence politically consequential. Borrow that institutional question while avoiding the claim that biology is fully predictable code.

Follow independent experimental confirmation, failure rates, performance beyond the training distribution, manufacturing consistency and the difference between a surrogate measurement and the desired outcome.

Brixi et al., Genome modelling and design across all domains of life with Evo 2 ; 4 March 2026 · Frances Arnold, Innovation by Evolution: Bringing New Chemistry to Life ; Nobel lecture

Useful quantum computing depends on corrected operations

Physical qubits are noisy. Quantum error correction spreads one protected unit of information across several physical qubits. Adding hardware helps only when errors, control and correction satisfy demanding conditions.

Why a machine with more physical qubits can be less useful than one with fewer but better controlled qubits. Memory, reliable logical operations and useful algorithms are distinct milestones.

What does not follow: A protected memory demonstration is not a complete practical computer. Qubit count alone does not establish cryptographic capability, commercial advantage or speedups for every AI task.

Willow’s below-threshold memory result appeared online in December 2024, in a 2025 Nature issue; its April 2026 correction amended figure labels. A separate 8 July 2026 paper reports reinforcement-learning control plus decoder steering improving logical stability 3.5-fold against injected drift; a specific test condition.

A Planetes-style maintenance perspective is the useful analogy, not a claim that Planetes depicts quantum error correction: the breakthrough still needs technicians, calibration and dependable supporting machinery.

Track logical error rates during actual operations, circuit depth, correction overhead, classical decoding, total runtime and independently reproduced advantage on a useful workload.

Google Quantum AI, Quantum error correction below the surface code threshold ; online December 2024; issue 2025 · Author correction ; April 2026 · Sivak et al., Reinforcement learning control of quantum error correction ; 8 July 2026

AI agents: six directions to follow

Try several plans, then check the result

An agent can spend extra computing time comparing possible answers, revising a draft or checking intermediate steps. The checker matters as much as the planner: repeating a mistake more carefully is still a mistake.

Snell and colleagues' August 2024 research found that allocating extra computation according to problem difficulty improved performance on tested reasoning problems. In some settings, this helped a smaller model compete with a larger one.

Not yet proven: This does not establish that longer reasoning always helps, that a model can reliably check itself, or that a plan will survive changing real-world conditions.

Independent checks, success per dollar, and whether added thinking helps on unfamiliar tasks rather than just familiar tests.

Ex Machina: the test itself can mislead the evaluator. A convincing explanation should not replace independent evidence.

Scaling LLM Test-Time Compute Optimally ; authors' paper (2024)

Learn what actions do to the world

A world model learns useful patterns from observations, then predicts what might happen after an action. A robot can compare possible moves before choosing one, without having to generate a realistic film of the future.

V-JEPA 2.1 tested this approach on a laboratory robot. Cup-grasping success was 70% with three-second planning and 80% with fourteen-second planning, averaged over ten tasks per skill. These are small experiments.

Not yet proven: Predicting useful visual features does not establish general physical understanding, reliable rescue work or safe behavior in arbitrary homes.

Unfamiliar objects, unexpected contact, recovery after mistakes and total human intervention time; not just successful demonstration clips.

Planetes: tools must work around fragile equipment and other people. Knowing when a prediction is unreliable becomes part of competence.

V-JEPA 2.1 ; authors' paper, Table 6 (2026)

Improve the learning rule itself

Most agents follow a learning method chosen by people. Researchers can also search for better learning methods by measuring how well many agents learn. This changes a controlled part of the system, within rules and tests supplied by researchers.

DiscoRL, published on 22 October 2025, learned update rules from agents playing games. A rule discovered on 57 Atari games transferred to other benchmarks, including ProcGen.

Not yet proven: This is not unrestricted self-rewriting or evidence that improvement will accelerate forever. The researchers still supplied the architecture, rewards, environments and search process.

Transfer to genuinely new tasks, repeatable gains after several improvement rounds, and checks that improvements preserve other useful skills.

A book premise: the most valuable asset is a difficult training environment that competitors cannot reproduce, rather than a single secret algorithm.

Discovering state-of-the-art reinforcement learning algorithms ; Nature (2025)

Propose discoveries that can actually be tested

A research agent becomes more useful when it can propose a candidate, run a meaningful test and use the result to improve its next attempt. The key resource is often a trustworthy test, not the number of agents debating.

AlphaEvolve's 2025 paper describes a system that changes program code and evaluates the results. The authors report new, verifiable algorithms in mathematics and computing, alongside improvements to selected computing workloads.

Not yet proven: Success where a computer can score an answer does not establish independent experimental science. A simulation can be wrong, and biological or materials claims still need suitable physical measurements.

Independently reproduced results, discoveries on problems chosen before testing, and the cost of failed experiments as well as successful ones.

The Windup Girl suggests a useful conflict: control over laboratories and production can matter more than access to promising designs.

AlphaEvolve: A coding agent for scientific and algorithmic discovery ; authors' paper (2025)

Finish the whole job, including recovery

A useful agent must carry a task through mistakes and changing information. A strong average score can hide particular tasks it almost always fails.

METR measures success on selected software tasks labeled by how long a human expert would need. Its dashboard distinguishes 50% and 80% success horizons. These describe task difficulty, not hours an AI can safely work alone.

Not yet proven: A longer measured horizon does not establish dependable performance across professions. The tasks are usually clearer and easier to score than ordinary work; extrapolating a growth curve is a forecast.

Repeated success on messy work, failure detection, recovery costs and performance when a knowledgeable person judges the whole result.

Patlabor 2: institutions need systems that remain usable during confusion, not merely impressive performance under a clean test procedure.

Task-Completion Time Horizons ; METR methodology and limitations; page marked updated 8 May 2026

Give agents limited authority and count the full cost

An agent needs clear limits on what it may spend, access or change. A useful delegation also has to save more work than its mistakes, supervision and computing bills create. Those are separate problems from answer quality.

Google's September 2025 Agent Payments Protocol proposes signed records of purchase instructions and approvals, including conditions for delegated purchases. This is concrete infrastructure for proving authority.

Not yet proven: A protocol announcement does not prove broad adoption, profitable autonomous businesses or wise purchases. A valid signature cannot decide whether the instructions served the user's interests.

Actual transaction use, refunds and disputes, review time, permission failures and total cost per successfully completed job.

Infomocracy and After Yang offer a useful question: who controls the records, rules and repair options when people depend on a service?

Announcing Agent Payments Protocol ; Google Cloud, 16 September 2025

The nonfiction reading guide

Algorithms to Live By

Brian Christian and Tom Griffiths · 2016 · Algorithms

How do you choose when time and information are limited?

Makes search, scheduling and trade-offs understandable through ordinary decisions. It gives you a better way to ask what an algorithm is optimizing.

Start with: The explore-versus-exploit discussion: when should you try something new?

Keep in mind: A tidy mathematical problem leaves out much of real life. The famous stopping rules depend on assumptions; they are not instructions for choosing a partner or a career.

Author or publisher

The Alignment Problem

Brian Christian · 2020 · AI

Why can a system do exactly what it was trained to do and still cause harm?

Connects training data, rewards and human judgment through reported research stories. Useful when someone promises that a better model will automatically be a better colleague.

Start with: The examples of biased data and badly chosen rewards.

Keep in mind: Published before the current wave of chat assistants. Read it for the problems and research history, then check newer evidence.

Author or publisher

Artificial Intelligence: A Guide for Thinking Humans

Melanie Mitchell · 2019 · AI

Does passing a test mean a machine understands the problem?

Explains why tasks that look similar to us can be very different for a machine. A useful foundation for checking claims about general intelligence.

Start with: The contrast between recognizing a pattern and grasping its meaning.

Keep in mind: The systems described are older. Some failures have improved; the test of whether skills transfer to unfamiliar situations still matters.

Author or publisher

Being You: A New Science of Consciousness

Anil Seth · 2021 · Mind

How does a brain build the experience of being a person?

Introduces perception as an active process: brains use expectations as well as signals from the world and body. Helps separate intelligence, awareness and a convincing imitation of a person.

Start with: The everyday perception examples before the wider claims about selfhood.

Keep in mind: This presents a research perspective, not a settled explanation of consciousness. It does not supply a test that can prove an AI has feelings.

Author or publisher

A Thousand Brains

Jeff Hawkins · 2021 · Mind

Could intelligence grow from many small models of the world working together?

Offers a concrete alternative to treating a brain as one giant predictor. Its emphasis on movement, reference frames and shared models is useful for thinking about robots.

Start with: The account of learning an object by touching and moving around it.

Keep in mind: Hawkins is arguing for his own theory and research program. It is not a complete, independently settled account of the brain.

Author or publisher

The Book of Why

Judea Pearl and Dana Mackenzie · 2018 · Reasoning

What would happen if we changed something, rather than only watched it?

Shows why a correlation cannot by itself tell you which action to take. Especially useful for evaluating job, health or economic forecasts that confuse prediction with cause.

Start with: The ladder from observing, to intervening, to asking what might have happened.

Keep in mind: A causal diagram contains assumptions. Drawing arrows does not make those assumptions true; evidence and subject knowledge still matter.

Author or publisher

The Beginning of Infinity

David Deutsch · 2011 · Discovery

What makes an explanation useful, and can discovery keep going?

A physicist argues that better explanations allow people to solve problems that once seemed permanent. Read it to challenge both fatalism and easy technological optimism.

Start with: The opening discussion of explanations that are hard to change without breaking them.

Keep in mind: Its claims about progress and knowledge are philosophical arguments, not a timetable or proof that every problem will be solved.

Author or publisher

Infinite Powers

Steven Strogatz · 2019 · Mathematics

How can dividing a problem into ever smaller pieces help us solve it?

Makes calculus approachable through its history and uses. It explains part of the mathematical toolkit behind motion, optimization and scientific models.

Start with: The early examples of approximating curves with simpler shapes.

Keep in mind: This is about calculus and its use of infinity. It is not a complete guide to set theory or evidence for physically infinite computing power.

Author or publisher

Quantum Computing Since Democritus

Scott Aaronson · 2013 · Quantum

Which problems can a computer solve, and what changes if it uses quantum physics?

Connects quantum computing to the deeper limits of computation. Useful protection against claims that quantum machines will simply try every answer and instantly solve everything.

Start with: The free lecture notes on sets, computation and quantum computing linked by the author.

Keep in mind: More demanding than a popular introduction. Hardware progress has changed since publication; the book is chiefly about ideas and limits.

Author or publisher

Bitcoin and Cryptocurrency Technologies

Arvind Narayanan, Joseph Bonneau, Edward Felten, Andrew Miller and Steven Goldfeder · 2016 · Web3

How can strangers agree on a shared record without one central operator?

Explains the machinery, security assumptions and privacy limits behind Bitcoin. The authors provide a legal, free pre-publication draft and teaching material.

Start with: The introduction to consensus and the section on anonymity.

Keep in mind: A foundation, not a guide to current tokens or every later blockchain. It gives no reason to assume that a useful protocol makes its token a good investment.

Author or publisher

The Blockchain and the New Architecture of Trust

Kevin Werbach · 2018 · Web3

If code replaces one intermediary, who do we end up trusting instead?

Looks at the people, rules and institutions around a blockchain. Useful for testing whether a proposed Web3 service solves a real coordination problem.

Start with: The distinction between trusting a system and trusting each person inside it.

Keep in mind: An argument about governance, not a current security audit or investment guide. Particular products and legal arrangements need fresh checks.

Author or publisher

Power and Progress

Daron Acemoglu and Simon Johnson · 2023 · Work

Who receives the benefits when technology makes work more productive?

Uses economic history to question the assumption that technical progress automatically raises everyone’s living standards. Helps frame career choices alongside workplace power and public policy.

Start with: The contrast between tools that help workers and systems designed mainly to replace them.

Keep in mind: The authors advance an interpretation and policy agenda. Historical comparisons inform choices; they do not give a formula for predicting a particular job.

Author or publisher

Chip War

Chris Miller · 2022 · Infrastructure

Why do a few factories and suppliers matter so much to world politics?

Traces how semiconductor production became both highly specialized and globally connected. Gives the physical background often missing from discussions of an AI race.

Start with: The passages on manufacturing specialization and dependence on difficult-to-replace suppliers.

Keep in mind: The industry and export rules change quickly. Use the book for history, and current company filings and government rules for today’s position.

Author or publisher

How the World Really Works

Vaclav Smil · 2022 · Infrastructure

What must be mined, made and powered before a digital future can exist?

Brings energy, food and materials back into forecasts that treat the economy as software. A useful prompt to ask about steel, cement, fertilizer and supply chains.

Start with: The chapters on energy and the materials that modern life depends on.

Keep in mind: Physical constraints matter, but this book’s judgments about the speed of change are arguments to test, not fixed laws of history.

Author or publisher

The Structure of Scientific Revolutions

Thomas S. Kuhn · 1962 · Discovery

Why do scientists sometimes disagree about what counts as a good answer?

Explores how research communities work within shared assumptions and what happens when those assumptions become strained. Helpful for reading claims of a revolutionary discovery.

Start with: The contrast between routine problem-solving and a change in the framework used to solve problems.

Keep in mind: An influential historical framework, not permission to call any unsupported idea a new paradigm. New claims still need evidence and replication.

Author or publisher

Reinforcement Learning: An Introduction, second edition

Richard S. Sutton and Andrew G. Barto · 2018 · Build

How does an agent learn which actions lead to a better result?

The technical foundation for learning through action and feedback. The authors make the book available online, so you can move from headlines to algorithms and exercises.

Start with: Chapter 1 and the multi-armed bandit examples before attempting the later mathematics.

Keep in mind: Requires more mathematics than the other starting points here. Learning in a defined environment does not establish reliable behavior in an open-ended workplace.

Author or publisher