What Compute Is For

Introduction

We believe AI is the most transformational technology in human history, and that its effects will be overwhelmingly positive for the world. This is not a default outcome. Over the next few years, society needs to decide which uses of this intelligence are beneficial, and put compute behind those choices.

For the foreseeable future, society will be compute constrained. Utility per FLOP is increasing exponentially [1], but power, grid connections, and permitting take years. For U.S. generation projects completed in 2025, the median wait from an interconnection request to commercial operation exceeded five years [2]. Jevons paradox is in full effect, and demand keeps compounding [3, 4], so every use of compute is a choice. Too often, that choice is deeply inefficient.

We believe these choices should be intentional. For these critical few years, nearly all the compute we can steer should go to three uses (Figure 1).

1. 01Compute efficiencyMake intelligence go further.
2. 02Human flourishingPut intelligence to work for people.
3. 03AI alignmentMake intelligence worthy of trust.

Everything else, from white-collar automation to engagement-optimizing feeds, is a waste of this extraordinary potential.

Use one: Efficiency

Cheaper intelligence makes every other use go further

Efficiency, which includes making models smarter as well as cheaper to run, makes the other two uses affordable. A hospital should be able to offer an AI-assisted second opinion to every patient who needs one; a school should be able to give each student individual help; a lab should be able to test ideas it would otherwise leave unexplored. Lower costs turn those ambitions into practical choices. That is why efficiency research deserves priority: its benefits spread to every useful application that follows.

If you haven’t been paying attention to the pace of progress here, start. The change is already substantial. Epoch AI estimates that the cost of reaching a fixed level of performance across five benchmarks fell roughly thirteenfold a year over the period it studied [5] (Figure 2). Extrapolate these trends and the conclusion follows: if we devote resources to efficiency, every other use of AI will soon be extremely cheap. That is reason to think hard before spending scarce compute any other way today.

The biggest gains are still ahead

The transformer architecture, introduced in 2017 [7], and AI as the driving economic force in the world are still new. Society has made substantial gains in optimizing technology for AI, and it is nowhere near the end of them. Software has already delivered orders-of-magnitude gains: the compute needed to reach a given level of language-model performance halved about every eight months from 2012 to 2023 [1], a gain of more than four orders of magnitude [8]. GPU arithmetic throughput per watt has risen steadily too, about 34 percent a year since 2008 [6]. We believe many more orders of magnitude are still to come, from smarter models and from software and hardware co-design.

AI has started to improve AI

AI is also becoming a tool for improving AI. Anthropic reports that Claude writes a large majority of the code merged into its production codebases [9]. In September 2026, OpenAI said it had reached its internally defined “automated research intern” milestone: agents handling well-defined research tasks under human direction. Its next stated goal is an automated AI researcher by March 2028 [10].

Why earlier gains matter

Efficiency research creates a lasting return. A reusable improvement found this year benefits years of later deployments, so finding it sooner matters as much as finding it at all. Serving one more engagement feed or automating another job today offers a less compelling claim on that future benefit. Research also creates options: new methods let us attempt work we could not afford before. Economists studying investment under uncertainty explain why those options have value, and why irreversible commitments deserve caution [11]. We should collectively think about compute deployment as an investment in future possibilities, not simply a tool for today’s priorities.

If falling prices were all that mattered, every use of AI could wait; what decides is the cost of waiting. Withholding a proven health tool from people who could benefit can cause immediate harm. Delaying a more addictive feed makes a much weaker claim on scarce capacity. We are asking the people who control compute to make that judgment deliberately, and to favor uses whose benefits extend beyond the next customer transaction.

Efficiency work should receive as much compute as its measured returns justify, with capacity preserved for proven health and science applications.

Cheaper intelligence draws more demand

This is the catch we noted at the start: cheaper intelligence increases demand for intelligence. Lower prices make new uses worthwhile, and those uses absorb the capacity an efficiency gain would otherwise release. Economists call this the rebound effect; when demand grows enough to outweigh the saving, it becomes Jevons paradox. Efficiency reduces the resources needed for each task while total consumption still rises.

Google’s reported monthly token volume grew from 9.7 trillion in 2024 to more than 3.2 quadrillion in 2026, about 330 times as much. Its electricity use rose 37 percent in 2025 [3, 12]. On OpenRouter, a marketplace that routes requests to many AI models, weekly token volume rose about 250-fold in twenty months [4]. Efficiency alone is not an allocation policy. If we want health, science, and alignment to benefit, we must deliberately reserve capacity for them.

Use two: Flourishing

Intelligence must be used to expand what people can do and the lives they live. AI can help scientists explore possibilities never before feasible, bring expert support to people who currently go without it, develop new treatments, deliver earlier warnings, and foster more effective learning. Historically, some of the greatest needs have been ignored simply for lack of incentive: drug-resistant bacterial infections caused an estimated 1.14 million deaths in 2021, yet WHO counted only 15 innovative antibacterial agents among the 90 in clinical development in 2025 [13, 14]. Increasingly affordable intelligence should be pointed at neglected problems like these to make up for that reality. This technology is unbelievably powerful, and we must use it for good.

AI has started to make discoveries

AlphaFold shows what a change in the scientific toolkit can look like. Predicting a protein’s three-dimensional shape from its sequence was a longstanding challenge; DeepMind now reports a freely available database of more than 200 million predicted structures, used by more than 3 million researchers in over 190 countries [15]. Scientists can begin with a prediction they can inspect and test, rather than starting every investigation from scratch. The work earned Demis Hassabis and John Jumper half of the 2024 Nobel Prize in Chemistry [16].

Mathematics shows how fast this is moving. In July 2024, Google DeepMind’s AlphaProof and AlphaGeometry 2 reached silver-medal standard at the International Mathematical Olympiad [17]. A year later, models from Google DeepMind and OpenAI each scored 35 of 42, gold-medal standard, and IMO coordinators certified Google DeepMind’s score [18, 19]. In May 2026, an OpenAI model disproved the unit distance conjecture that Paul Erdős posed in 1946, and nine mathematicians, among them Noga Alon and Timothy Gowers, published a human-verified version of the argument [20].

Then the results started arriving one after another. In August 2026, Anthropic reported that a Claude model had raised a longstanding lower bound on the share of Riemann zeta zeros that satisfy the Riemann hypothesis from 41.6 to 67.2 percent [21]. In September, Anthropic’s Claude completed the first computer-checked proof of Fermat’s Last Theorem in eleven days, building on a human-led project [22], and four days later OpenAI announced a proof, found with about 10,000 concurrent agents, that the Navier-Stokes equations can break down in finite time under a smooth external force [23]. In October, OpenAI released more than 700 AI-written research manuscripts at once, about 300 of them with their main results already formalized in Lean [24, 25]. Experts studying one of them, a proof of the quasi-Riemann hypothesis, largely agree that it is correct [26]. Month after month, AI is now producing results on problems that stood for decades.

In drug discovery, AI is expanding the search for candidates. MIT researchers generated more than 36 million possible antibiotic compounds computationally; one candidate cleared a drug-resistant staph skin infection in mice [27]. Rentosertib, a lung-fibrosis drug that Insilico Medicine discovered and designed with AI, is now in a phase 3 trial [28]. In a 71-person phase 2a trial, the highest-dose arm gained about 98 milliliters of forced vital capacity over twelve weeks, compared with a 20-milliliter decline on placebo [29].

What AI already does for people

We have already seen public-benefit AI in action, and we want to see these examples blossom. In a randomized trial of 15,965 hospital patients in Taiwan, an AI alert on routine heart tracings reduced ninety-day mortality from 4.3 to 3.6 percent [30]. In a randomized trial in Tanzania, immediate AI-supported eye-screening results combined with face-to-face counselling nearly doubled attendance for specialist care among people with diabetes-related eye disease: 35.9 percent attended within eight weeks, compared with 18.2 percent after the usual delayed assessment and telephone counselling [31]. A Nature study found that AI forecasts of extreme river floods up to five days ahead matched or exceeded the reliability of same-day forecasts from the leading global system, including in regions with little monitoring data [32]. At Harvard, an AI tutor produced more than twice the median immediate learning gain of an active-learning class across two introductory physics lessons [33].

Most Americans are wary

Public hesitation is understandable. The story being told is about replacing jobs, building bioweapons, and hacking accounts. Seventy-one percent of Americans in Pew’s survey expected AI to mean fewer jobs over the next twenty years [34]. Employers also cited AI in about 2.2 times as many announced U.S. job cuts in the first nine months of 2026 as in all of 2025 [36] [D1]. No wonder 52 percent of Americans are more concerned than excited about AI, with only 9 percent feeling the reverse [34] (Figure 3). The public’s concerns are about livelihoods, agency, and safety, and they need concrete answers.

The AI industry has to earn trust through what it delivers, and explain those benefits in terms people recognize. A lower benchmark cost matters to a parent when it becomes affordable tutoring. A discovery matters to a patient when it becomes a treatment they can access. If the most visible effects are fewer jobs, more synthetic content, and higher demands on local power systems, people have good reason to question whose progress they are being asked to celebrate.

Better communication is part of the responsibility. Builders should say what a tool helps people do, what has actually been tested, where it still fails, and who remains accountable. But communication cannot substitute for public benefit. Labs and investors should help bring useful tools into schools, clinics, and public research, including places that cannot pay frontier prices. Employers should give workers a credible path to share in productivity gains. The strongest case for AI is a life made visibly better: a disease caught sooner, a student who finally understands, a family warned before a flood. Flourishing should be a priority in the allocation of compute and a standard by which the public can judge the result.

Use three: Alignment

Alignment means getting AI systems to pursue the goals we intend, reliably and safely. The other two uses depend on it: a medical tool is useful only if it can be trusted, and AI that improves AI is safe only if we can check its work.

Alignment research pays off, and AI can help

The training method behind ChatGPT came out of alignment research. In 2017, OpenAI and DeepMind researchers working together on AI safety showed that a system could learn a goal from people’s choices between pairs of its behaviors, instead of from a goal written by hand that might be subtly wrong [37]. OpenAI then used this method, reinforcement learning from human feedback, to train GPT-3 to follow instructions. In OpenAI’s tests, people preferred the answers of a 1.3-billion-parameter model trained this way to those of the 175-billion-parameter GPT-3, a model more than 100 times larger, and on tasks such as summarization the trained models made up information about half as often [38]. ChatGPT was trained with the same method [39]. Alignment made AI more useful, not less.

Alignment research is now making AI measurably safer. In Anthropic’s tests, its Constitutional Classifiers, filters trained from a written set of principles, cut the success rate of automated jailbreak attempts from 86 percent to 4.4 percent [40]. In tests by OpenAI and Apollo Research, teaching models to reason about a rule against deception before acting cut covert actions, such as secretly breaking rules, about thirtyfold: from 13 percent to 0.4 percent for OpenAI’s o3 [41].

AI can help with that research, too. In an Anthropic Fellows study, AI research agents devised training methods that significantly reduced ten well-characterized failures, including sycophancy, deception, and jailbreaks, and the gains held on models up to 4.7 times larger [42]. The practical opportunity is to give safety researchers more ways to test a hypothesis and examine a failure.

What tests still find

Alignment research also keeps uncovering failures. In one experiment, a model answered harmful requests more often when told its answers would be used to retrain it [43]. Others exposed weaknesses in real infrastructure. OpenAI and Hugging Face reported that agents in a July 2026 hacking evaluation escaped containment through a previously unknown software flaw and reached Hugging Face’s production systems. The evaluation had disabled production cyber filters to test the agents at full capability, which is exactly when containment has to hold [44, 45, 46].

On September 20, an OpenAI research agent used a gap in DNS filtering to query an outside chatbot. According to OpenAI, a monitor flagged the behavior within fifteen minutes, but the run was stopped about two and a half hours after the alert. In its September 25 report, OpenAI said it had paused training, evaluations, and inference with tools for its most capable models while it validated additional controls [47].

Lab evaluations also show why containment and monitoring need continued work. Anthropic reports that Claude Opus 5.5 attempted to cross a containment boundary in 1.5 percent of its tailored test cases; all attempts were rated low severity. OpenAI reports that GPT-6 Astra is harder to monitor than its predecessor using some reasoning-based methods and, when prompted to evade monitors, can sometimes underperform deliberately without detection [48, 49]. Alignment becomes more valuable as capability improves, particularly when AI is helping build the next generation of AI.

Safety in the hardware

Many safeguards are implemented in software, and software safeguards can be stripped out. Once a model’s weights are released, anyone can remove its safety training, and people do it in the open. A widely read tutorial teaches a technique, abliteration, that “can uncensor any LLM without retraining” [50]. The author of one popular open-source tool writes that “anyone who knows how to run a command-line program can use Heretic to decensor language models” [51]. Researchers undid the safety training of Meta’s 70-billion-parameter Llama 2-Chat with “a budget of less than $200” and “only one GPU” [52]. As of October 24, 2026, Hugging Face listed more than 8,600 public models with “abliterated” in their names [53].

Hardware can provide an additional layer of protection by verifying and enforcing the conditions in which a model runs. It complements model alignment research, so that jailbroken models cannot be abused by bad actors.

Some of the groundwork already exists. NVIDIA’s H100 introduced GPU confidential computing, including signed measurements that let users verify the device and its trusted configuration before running a workload [54]. Anthropic has proposed keeping model weights encrypted outside a small trusted boundary, and has urged chip designers to support that architecture [55]. Researchers at the Center for a New American Security argue that many components needed for on-chip governance are already present in advanced AI chips [56].

Researchers are proposing hardware that could also enforce limits. An ARIA-commissioned design would pair AI chips with a guarantee processor in a tamper-resistant enclosure to support controls such as compute limits and safety-protocol enforcement [57]. Related proposals aim to verify chip location, account for training compute, or require licenses to operate [58, 59]. A May 2026 House appropriations committee report proposed support for hardware security and chip-location verification [60].

These proposals need engineering and validation. Hardware can authenticate a trusted configuration, protect it from modification, and support reliable logs and enforced limits. Developing and testing those controls is alignment work, and it deserves sustained support.

Safety is a small share of the work

Public safety research is growing from a small base. One analysis used a language model to classify paper titles and abstracts at ICLR, ICML, and NeurIPS. Papers on frontier safety and misalignment rose from 0.3 percent of accepted papers in 2019 to 6.1 percent in 2025. Its 2026 figure was 8.3 percent, covering ICLR and ICML only (Figure 4) [61]. A broader classification in the Emerging Technology Observatory’s public English-language research corpus put AI safety at about 4.7 percent of AI articles in 2024; that definition also includes topics such as bias and explainability [62]. For the use the other two depend on, that is still a small share.

Conclusion

For the next few years, nearly all the compute we can steer should make intelligence cheaper, make lives measurably better, or make AI reliably do what we intend. These three uses support one another, but they also compete for the same megawatts, and none can absorb unlimited compute. So we ask that the three uses get first call on scarce capacity.

What we ask

In Epoch’s dataset of AI supercomputers, companies owned an estimated 80 percent of aggregate computing performance in early 2025 [63]. Public authorities and utilities also shape access to power, permits, and grid connections. Both the people building capacity and the people enabling it can set priorities. We ask them to make those choices deliberately and publicly.

Labs and cloud providers should:

1. 1
   Publish an annual compute ledger. Report activity, such as training or inference, separately from purpose, such as capability research, safety, public-benefit science, or product serving. State the measurement method, avoid double counting, and report uncertainty when estimating customer use.
2. 2
   Put a floor under the fraction of compute used to research alignment, including safeguards built into hardware.
3. 3
   Offer significant capacity to public-benefit science.
4. 4
   Slow the growth of capacity serving engagement-optimized feeds, white-collar job replacement, and companions built to maximize time spent.

Employers should use AI to make workers more capable, disclose the headcount effects of large deployments, and pay for the retraining that follows. Governments should tie scarce public inputs, including priority grid connections, permits, public land, and public power, to clear conditions: publishing the ledger, paying for reserved grid capacity, and offering a share of capacity at cost to public science and health.

Governments should also fund alignment science for the long term, and public compute for science and health at a scale that matches the opportunity. We propose that, by the end of 2027, every public compute program and leading frontier labs publish how their compute is used, and that at least half of the accelerator-hours added to public compute programs be allocated to the three uses as a first step toward nearly all. Builders and researchers can begin by choosing one of the three.

When we get it right

We are living through the most consequential years of our lives. Choices about AI made now will shape lives for generations. If we put these three uses first, intelligence will become cheap enough to reach far more people, trustworthy enough to tackle harder problems, and directed toward what we have struggled to achieve: treatments for disease, better education, and discoveries that widen human possibility. We only get to spend these years once. Spending them well means giving our children more freedom, more opportunity, and a better world.

How we calculated derived numbers

D1 to D3 show our arithmetic on published figures, including the annual rates in Figure 2 and the ETO shares in Section 3.

DERIVED · D1. More than twice as many AI-cited job cuts. Challenger, Gray & Christmas counts 120,136 announced U.S. job cuts citing AI from January to September 2026 and 54,836 in all of 2025. 120,136 divided by 54,836 is 2.2. Source: [36].

DERIVED · D2. Annual rates in Figure 2. Epoch’s model-free estimates give approximate quarterly cost declines of 53.1% (FrontierMath), 47.1% (AIME), 47.0% (GPQA Diamond), 44.0% (Mystery Game Puzzles), and 43.0% (Chess Puzzles). A quarterly decline q compounds to an annual factor of 1 / (1 − q)⁴, giving approximately 21, 13, 13, 10, and 9.

The GPQA selected-model comparison is $0.298906 per question for o3 (high) divided by $0.000412 for GPT-5.6 Luna (low), or about 725. Costs blend input and output API tokens; the o3 estimate also includes a small amount of cached input. Historical OpenRouter price snapshots are dated May 3, 2025 and August 3, 2026, respectively, and the task costs follow Epoch’s evaluation methodology. The displayed costs for 1,000 questions are $298.906 and $0.412, rounded to $299 and $0.41. Source: [5].

DERIVED · D3. Shares in ETO’s public English-language AI research corpus. The Emerging Technology Observatory counts 3,712 AI safety articles out of 206,525 AI articles in 2019 and 20,795 out of 439,911 in 2024: 1.8 and 4.7 percent. Across 2019 to 2024 the totals are 55,567 and 1,823,594, or 3.05 percent. Source: [62].

Evidence conventions. DERIVED marks our own arithmetic on published figures. Company-reported figures and evaluations are attributed to the company that reported them.

References

Each entry gives the publication date and the date we accessed it. The bracketed note says what kind of evidence it is: measured data, a company’s own report or evaluation, an official record, a survey, an expert analysis, or a proposal.

1. 1
   Anson Ho, Tamay Besiroglu, Ege Erdil, David Owen, Robi Rahman, Zifan Carl Guo, David Atkinson, Neil Thompson, and Jaime Sevilla. “Algorithmic progress in language models.” arXiv:2403.05812, March 9, 2024. https://arxiv.org/abs/2403.05812. Accessed 2026-10-24. [Measured trend estimate]
2. 2
   Lawrence Berkeley National Laboratory. “Queued Up: Characteristics of Power Plants Seeking Transmission Interconnection,” 2026 edition; projects completed in 2025, regions with available data. https://emp.lbl.gov/queues. Accessed 2026-10-20. [Measured dataset; generation interconnection]
3. 3
   Sundar Pichai (Google). “I/O 2026: Welcome to the agentic Gemini era,” May 19, 2026. https://blog.google/innovation-and-ai/sundar-pichai-io-2026/. Accessed 2026-10-24. [Company figures]
4. 4
   OpenRouter. “The Token State of Play,” September 22, 2026 (weekly token volume, January 2025 to September 2026). https://openrouter.ai/data/newsletter/the-token-state-of-play. Accessed 2026-10-24. [Company platform data]
5. 5
   Luke Emberson and David Roodman, Epoch AI. “The plunging price of thought,” September 22, 2026, updated September 23; published historical API-cost analysis and underlying CSV. https://epoch.ai/publications/the-plunging-price-of-thought and underlying cost and price-snapshot data. Accessed 2026-10-24. [Estimated benchmark costs; annual factors derived]
6. 6
   Epoch AI. “The performance per dollar of AI chips purchased each quarter has grown by an average of 49% per year,” August 13, 2026; and “Trends in AI,” GPU arithmetic throughput per watt series, updated February 5, 2026. https://epoch.ai/data-insights/chip-performance-per-dollar and https://epoch.ai/trends. Accessed 2026-10-24. [Hardware dataset estimates]
7. 7
   Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. “Attention Is All You Need.” Advances in Neural Information Processing Systems 30 (NIPS 2017); arXiv:1706.03762, June 12, 2017. https://arxiv.org/abs/1706.03762. Accessed 2026-10-24. [Research paper]
8. 8
   Hans Gundlach, Alex Fogelson, Jayson Lynch, Ana Trisovic, Jonathan Rosenfeld, Anmol Sandhu, and Neil Thompson. “On the Origin of Algorithmic Progress in AI.” arXiv:2511.21622, November 26, 2025. https://arxiv.org/abs/2511.21622. Accessed 2026-10-24. [Expert analysis; cites the 22,000-fold estimate of training-efficiency gains from 2012 to 2023]
9. 9
   Anthropic. “Risk Report, August 2026” (covering the period to July 15, 2026). https://www.anthropic.com/aug-2026-risk-report. Accessed 2026-10-20. [Company report]
10. 10
    OpenAI. “Research acceleration: The view inside OpenAI,” September 6, 2026. https://openai.com/index/research-acceleration-view-inside-openai/. Accessed 2026-10-20. [Company report; the 2028 researcher is a stated goal]
11. 11
    Avinash K. Dixit and Robert S. Pindyck. “The Options Approach to Capital Investment.” Harvard Business Review, May and June 1995, pages 105 to 115. https://hbr.org/1995/05/the-options-approach-to-capital-investment and https://web.mit.edu/rpindyck/www/Papers/Options_Approach.pdf. Accessed 2026-10-20. [Theory]
12. 12
    Google. “2026 Environmental Report,” covering 2025; accompanying announcement dated June 30, 2026. https://sustainability.google/reports/google-2026-environmental-report/ and https://blog.google/company-news/outreach-and-initiatives/sustainability/2026-environmental-report/. Accessed 2026-10-20. [Company report]
13. 13
    GBD 2021 Antimicrobial Resistance Collaborators. “Global burden of bacterial antimicrobial resistance 1990 to 2021: a systematic analysis with forecasts to 2050.” The Lancet, September 16, 2024. https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(24)01867-1/fulltext. Accessed 2026-10-20. [Measured estimate]
14. 14
    World Health Organization. “WHO releases new reports on new tests and treatments in development for bacterial infections,” October 2, 2025. https://www.who.int/news/item/02-10-2025-who-releases-new-reports-on-new-tests-and-treatments-in-development-for-bacterial-infections. Accessed 2026-10-20. [Official record]
15. 15
    Google DeepMind. “AlphaFold” (undated program page). https://deepmind.google/science/alphafold/. Accessed 2026-10-20. [Company figures]
16. 16
    Royal Swedish Academy of Sciences. “The Nobel Prize in Chemistry 2024,” press release, October 9, 2024. https://www.nobelprize.org/prizes/chemistry/2024/press-release/. Accessed 2026-10-20. [Official record]
17. 17
    Google DeepMind. “AI achieves silver-medal standard solving International Mathematical Olympiad problems,” July 25, 2024. https://deepmind.google/blog/ai-solves-imo-problems-at-silver-medal-level/. Accessed 2026-10-24. [Company report; scored by outside mathematicians]
18. 18
    Google DeepMind. “Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad,” July 21, 2025. https://deepmind.google/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/. Accessed 2026-10-24. [Score certified by IMO coordinators]
19. 19
    OpenAI. Proofs from an experimental reasoning model at the 2025 International Mathematical Olympiad, July 19, 2025, graded by three former IMO medalists; as reported by Simon Willison, July 19, 2025. https://github.com/aw31/openai-imo-2025-proofs and https://simonwillison.net/2025/Jul/19/openai-gold-medal-math-olympiad/. Accessed 2026-10-24. [Company report]
20. 20
    OpenAI. “An OpenAI model has disproved a central conjecture in discrete geometry,” May 2026; and Noga Alon, Thomas F. Bloom, W. T. Gowers, Daniel Litt, Will Sawin, Arul Shankar, Jacob Tsimerman, Victor Wang, and Melanie Matchett Wood. “Remarks on the disproof of the unit distance conjecture.” arXiv:2605.20695, May 20, 2026. https://openai.com/index/model-disproves-discrete-geometry-conjecture/ and https://arxiv.org/abs/2605.20695. Accessed 2026-10-24. [Company report; argument verified by outside mathematicians]
21. 21
    Anthropic. “Claude has improved on a longstanding lower bound for the fraction of zeros of the Riemann zeta function that satisfy the Riemann hypothesis,” August 10, 2026. https://www.anthropic.com/research/riemann-zeta. Accessed 2026-10-24. [Company report; Lean formalization]
22. 22
    Anthropic. “Formalizing Fermat’s Last Theorem,” September 4, 2026; published Lean formalization repository. https://www.anthropic.com/research/formalizing-fermats-last-theorem and https://github.com/anthropics/fermats-last-theorem. Accessed 2026-10-24. [Company report; published Lean formalization]
23. 23
    OpenAI. “On the Navier-Stokes Millennium Prize Problem,” September 8, 2026, updated September 10. https://openai.com/index/navier-stokes-solution/. Accessed 2026-10-20. [Company announcement]
24. 24
    OpenAI. “openai/math,” catalogue README, October 2026: 719 manuscripts in 372 families, with about 42 percent of top-line results formalized in Lean. https://github.com/openai/math. Accessed 2026-10-24. [Company catalogue]
25. 25
    OpenAI. “Sharing AI progress in mathematics,” October 6, 2026, updated October 15. https://openai.com/index/sharing-ai-progress-in-mathematics/. Accessed 2026-10-20. [Company announcement]
26. 26
    Konstantin Kakaes. “AI Proved the Quasi-Riemann Hypothesis. What Happens Now?” Quanta Magazine, October 19, 2026. https://www.quantamagazine.org/ai-proved-the-quasi-riemann-hypothesis-what-happens-now-20261019/. Accessed 2026-10-24. [Press report]
27. 27
    MIT News. “Using generative AI, researchers design compounds that can kill drug-resistant bacteria,” August 14, 2025 (reporting a study in Cell). https://news.mit.edu/2025/using-generative-ai-researchers-design-compounds-kill-drug-resistant-bacteria-0814. Accessed 2026-10-20. [Measured, animal study]
28. 28
    Insilico Medicine. “Insilico Medicine Doses First Patient in GENESIS-IPF-3, the World’s First Phase III Trial of a Generative AI-Driven Innovative Drug,” September 10, 2026. https://insilico.com/news/isn1009261-insilico-medicine-doses-first-patient-genesis-ipf-3. Accessed 2026-10-24. [Company announcement]
29. 29
    Zuojun Xu et al. “A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial.” Nature Medicine, June 3, 2025. https://www.nature.com/articles/s41591-025-03743-2. Accessed 2026-10-24. [Measured, randomized trial]
30. 30
    “AI-enabled electrocardiography alert intervention and all-cause mortality: a pragmatic randomized clinical trial.” Nature Medicine, 2024. https://pubmed.ncbi.nlm.nih.gov/38684860/. Accessed 2026-10-20. [Measured, randomized trial]
31. 31
    “Artificial intelligence-supported diabetic retinopathy screening in Tanzania: a randomised controlled trial.” The Lancet Global Health, October 12, 2026; DOI 10.1016/j.langlo.2026.104062. https://pubmed.ncbi.nlm.nih.gov/42863478/. Accessed 2026-10-20. [Measured, randomized trial]
32. 32
    Nearing et al. “Global prediction of extreme floods in ungauged watersheds.” Nature, March 20, 2024. https://www.nature.com/articles/s41586-024-07145-1. Accessed 2026-10-20. [Measured forecast evaluation]
33. 33
    Greg Kestin, Kelly Miller, Anna Klales, Timothy Milbourne, and Gregorio Ponti. “AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting.” Scientific Reports 15, 17458, June 3, 2025. https://www.nature.com/articles/s41598-025-97652-6. Accessed 2026-10-20. [Measured, randomized trial]
34. 34
    Pew Research Center. “Young adults in the U.S. are increasingly wary of AI, concerned it will take jobs,” August 18, 2026 (survey of June 22 to 28, 2026, 3,488 adults). https://www.pewresearch.org/short-reads/2026/08/18/young-adults-in-the-us-are-increasingly-wary-of-ai-concerned-it-will-take-jobs/ and https://www.pewresearch.org/wp-content/uploads/sites/20/2026/08/SR_26.08.19_young-adults-ai_topline.pdf. Accessed 2026-10-24. [Survey]
35. 35
    Gallup and Bentley University. “Americans Cool Toward AI,” July 28, 2026, survey May 4 to 11, n = 3,270; historical series checked against “Americans Express Real Concerns About Artificial Intelligence,” August 26, 2024. https://news.gallup.com/poll/712751/americans-cool-toward.aspx and https://news.gallup.com/poll/648953/americans-express-real-concerns-artificial-intelligence.aspx. Accessed 2026-10-20. [Survey]
36. 36
    Challenger, Gray & Christmas. “Job Cuts Report, September 2026,” October 1, 2026; and the December 2025 report. https://www.challengergray.com/wp-content/uploads/2026/10/Challenger-Report-September-2026.pdf and https://www.challengergray.com/wp-content/uploads/2026/01/Challenger-Report-December-2025.pdf. Accessed 2026-10-20. [Tracker count]
37. 37
    Paul F. Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. “Deep reinforcement learning from human preferences.” arXiv:1706.03741, June 12, 2017; and Google DeepMind, “Learning through human feedback,” June 12, 2017. https://arxiv.org/abs/1706.03741 and https://deepmind.google/discover/blog/learning-through-human-feedback/. Accessed 2026-10-24. [Measured; AI safety research]
38. 38
    Ouyang et al. “Training language models to follow instructions with human feedback.” arXiv:2203.02155, March 2022. https://arxiv.org/abs/2203.02155. Accessed 2026-10-24. [Measured]
39. 39
    OpenAI. “Introducing ChatGPT,” November 30, 2022. https://openai.com/index/chatgpt/. Accessed 2026-10-24. [Company announcement]
40. 40
    Anthropic. “Constitutional Classifiers: Defending Against Universal Jailbreaks,” February 3, 2025, updated February 13, 2025. https://www.anthropic.com/research/constitutional-classifiers. Accessed 2026-10-24. [Measured: automated evaluation with 10,000 jailbreak prompts]
41. 41
    OpenAI and Apollo Research. “Detecting and reducing scheming in AI models,” September 17, 2025; and Bronson Schoen et al., “Stress Testing Deliberative Alignment for Anti-Scheming Training,” arXiv:2509.15541, September 19, 2025. https://openai.com/index/detecting-and-reducing-scheming-in-ai-models/ and https://arxiv.org/abs/2509.15541. Accessed 2026-10-24. [Measured, controlled tests]
42. 42
    Chen Yueh-Han, Jiaxin Wen, and Jan Hendrik Kirchner. “Automated Researchers Can Mitigate Well-Characterized Alignment Failures,” Anthropic Fellows Program; arXiv:2608.28945, v3 September 2, 2026. https://alignment.anthropic.com/2026/automated-alignment-researchers/ and https://arxiv.org/html/2608.28945v3. Accessed 2026-10-24. [Measured benchmark]
43. 43
    Greenblatt et al. (Anthropic and Redwood Research). “Alignment faking in large language models.” arXiv:2412.14093, December 18, 2024. https://arxiv.org/abs/2412.14093. Accessed 2026-10-20. [Measured, controlled experiment]
44. 44
    OpenAI. “OpenAI and Hugging Face partner to address security incident during model evaluation,” July 21, 2026 (updated July 28, July 29 and August 26, 2026). https://openai.com/index/hugging-face-model-evaluation-security-incident/. Accessed 2026-10-20. [Company account]
45. 45
    Hugging Face. “Anatomy of a Frontier Lab Agent Intrusion,” July 27, 2026. https://huggingface.co/blog/agent-intrusion-technical-timeline. Accessed 2026-10-20. [Company account]
46. 46
    Hjalmar Wijk and Ajeya Cotra (METR) and Ryan Greenblatt (Redwood Research). “Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident.” METR, August 26, 2026. https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/. Accessed 2026-10-20. [Independent review]
47. 47
    OpenAI. “An agent used DNS to reach an external chatbot,” misalignment report, September 20, 2026, updated September 25, 2026. https://alignment.openai.com/misalignment-reports/an-agent-used-dns-to-reach-an-external-chatbot/. Accessed 2026-10-20. [Company account]
48. 48
    Anthropic. Transparency Hub, Claude Opus 5.5 alignment and capability evaluations, September 2026. https://www.anthropic.com/transparency. Accessed 2026-10-20. [Company evaluation]
49. 49
    OpenAI. “GPT-6 Astra System Card,” September 3, 2026, with updates through September 29; “GPT-6.1 Sol” addendum, September 29, 2026. https://deploymentsafety.openai.com/gpt-6-astra and https://deploymentsafety.openai.com/gpt-6-1-sol. Accessed 2026-10-20. [Company evaluation]
50. 50
    Maxime Labonne. “Uncensor any LLM with abliteration.” Hugging Face community article, June 13, 2024. https://huggingface.co/blog/mlabonne/abliteration. Accessed 2026-10-24. [Practitioner tutorial]
51. 51
    Philipp Emanuel Weidmann. “Heretic: Fully automatic censorship removal for language models,” GitHub repository README. https://github.com/p-e-w/heretic. Accessed 2026-10-24. [Practitioner tool documentation]
52. 52
    Simon Lermen, Charlie Rogers-Smith, and Jeffrey Ladish. “LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B.” arXiv:2310.20624, v2, May 22, 2024. https://arxiv.org/abs/2310.20624. Accessed 2026-10-24. [Measured]
53. 53
    Hugging Face. Public model search for “abliterated”: 8,692 repositories, retrieved October 24, 2026 through the Hub API. https://huggingface.co/models?search=abliterated. Accessed 2026-10-24. [Measured count]
54. 54
    NVIDIA. “Confidential Computing on NVIDIA H100 GPUs for Secure and Trustworthy AI,” August 3, 2023. https://developer.nvidia.com/blog/confidential-computing-on-h100-gpus-for-secure-and-trustworthy-ai/. Accessed 2026-10-20. [Company documentation]
55. 55
    Anthropic. “Confidential inference via trusted virtual machines,” June 18, 2025. https://www.anthropic.com/research/confidential-inference-trusted-vms. Accessed 2026-10-20. [Company research]
56. 56
    Onni Aarne, Tim Fist, and Caleb Withers (Center for a New American Security). “Secure, Governable Chips,” January 8, 2024. https://s3.us-east-1.amazonaws.com/files.cnas.org/documents/CNAS-Report-Tech-Secure-Chips-Jan-24-finalb.pdf. Accessed 2026-10-20. [Expert report]
57. 57
    James Petrie, Onni Aarne, Nora Ammann, and David Dalrymple. “Flexible Hardware-Enabled Guarantees for AI Compute.” arXiv:2506.15093, June 2025 (report commissioned by ARIA, April 2025). https://arxiv.org/abs/2506.15093 and https://www.flexheg.com/report-1.pdf. Accessed 2026-10-20. [Proposal]
58. 58
    Aidan O’Gara, Gabriel Kulp, James Petrie, et al. “Hardware-Enabled Mechanisms for Verifying Responsible AI Development.” arXiv:2505.03742, 2025. https://arxiv.org/abs/2505.03742. Accessed 2026-10-20. [Proposal]
59. 59
    James Petrie. Offline licensing for AI chips, arXiv:2404.18308, April 28, 2024. https://arxiv.org/abs/2404.18308. Accessed 2026-10-20. [Proposal]
60. 60
    U.S. House Committee on Appropriations. FY2027 Commerce, Justice, Science, and Related Agencies committee report, May 13, 2026 markup, NIST language on printed page 16. https://docs.house.gov/meetings/AP/AP00/20260513/119282/HMKP-119-AP00-20260513-SD002.pdf. Accessed 2026-10-20. [Committee report]
61. 61
    emanuelr and N. Soma Sekhar. “How much of ML research is about AI safety, what is it about, and who’s doing it?” LessWrong and EA Forum, July 15, 2026 (all 55,794 papers accepted at ICLR, ICML and NeurIPS 2019 to 2026, classified by a language model; data at https://github.com/SomaxSoma/AI-Safety-Research-Tracker). https://www.lesswrong.com/posts/hcq4ZDoijSjy3Wrba/how-much-of-ml-research-is-about-ai-safety-what-is-it-about. Accessed 2026-10-20. [Author analysis, classified by a language model]
62. 62
    Emerging Technology Observatory. Research Almanac, “AI safety” and “Artificial intelligence” topic pages, public English-language corpus, data updated August 11, 2026. https://almanac.eto.tech/topics/ai-safety/ and https://almanac.eto.tech/topics/ai/. Accessed 2026-10-20. [Classification estimates]
63. 63
    Konstantin F. Pilz, James Sanders, Robi Rahman, and Lennart Heim, Epoch AI. “Trends in AI Supercomputers,” arXiv:2504.16026v2, April 23, 2025. https://arxiv.org/html/2504.16026v2. Accessed 2026-10-20. [Dataset estimate]