A lab says AI has made its research 100× faster

Where things stand (October 2026)

Nothing published since the question opened on 2026-09-16 comes close to qualifying. The labs that measure overall research pace all report small numbers:

• Anthropic, Risk Report, August 2026 (covers work through 15 July 2026). It says its internal AI research is "significantly faster than [it] would be without AI assistance, but not yet by a factor of 2". The method compares Anthropic's internal capability index (its version of Epoch's capability index) against the pre-AI trend and attributes the difference. Its rulebook trigger is a doubling of the rate of progress. The report calls 10^3–10^10× "effective scaleup within a year" the acute scenario.
• Anthropic, 17 September 2026 (inside the window). The post on its new R&D Automation Index says Claude "leads" 26% of research work and runs fully autonomously on none of it. That is a share of work, not a pace ratio.
• Anthropic, "When AI builds itself". Engineers merge about 8× more code than before, and researchers self-report about 4× output. Anthropic stresses that bottlenecks (Amdahl's law) keep the overall pace well below these task-level numbers.
• Google DeepMind, Gemini 4 Argon safety report, October 2026. It finds "no significant end-to-end acceleration". Its capability index rose about 0.091 points per month before April 2026 and about 0.118 after; the difference is not statistically significant.
• OpenAI, 6 September 2026 (before the window). It reports 3.1 agent-workdays per human workday and says it reached its "automated research intern" goal. It warns that overall pace "likely won't keep pace with these specific metrics". Its target is an automated AI researcher by March 2028.

So the measured overall speed-up is roughly 1–2×. A YES needs at least 100×, which is about two orders of magnitude more, plus a published method that meets strict criteria.

Reference class and model inputs

No lab has ever published an overall-pace ratio above 2×. Each lab's own threshold sits at 2–5×: Anthropic's rulebook trigger at 2×, OpenAI's "Critical" self-improvement level at 5×, DeepMind's "substantial acceleration" threshold. So 100× is far beyond every published threshold.

I computed takeoff probabilities from the AI Futures Model's August 2026 Monte Carlo results, pulled from aifuturesmodel.com/forecast/{daniel,eli,brendan}-08-16-26. These forecasters are at the aggressive end, and their numbers assume no deliberate slowdown.

• Chance of a superhuman AI researcher by 2032: 0.77 (Daniel), 0.57 (Eli), 0.72 (Brendan).
• Chance of a superintelligent AI researcher by 2032: 0.65, 0.43, 0.60.
• In the model's default run, a superhuman researcher means about a 17× research speed-up and a superintelligent one about 448×, so 100× falls between them.

Forethought's model (Davidson and Houlden) is more cautious. It gives about a 20% chance that full automation compresses more than 10 years of progress into one year, and says the pace "won't be 30X faster [over a year] absent a paradigm shift". The cap comes from limits on how far software can improve and from compute bottlenecks.

The September 2026 paper on an intelligence explosion (Hinton, Bengio, Pachocki and others) estimates about a tenfold speed-up within roughly 1.5 years after full automation.

Kalshi's thinly traded market on OpenAI or Anthropic announcing recursive self-improvement stands at about 19% before 2027, 48% before 2028 and 66% before 2029. That is a much lower bar than 100× with a method.

Paths to YES and NO

YES needs all of the following:

• Full automation of AI research at a lab that publishes, which in this fast-moving world is plausibly 2027–2029. Jack Clark now expects full self-improvement could be possible in 2027.
• A fast takeoff that pushes past 100× before bottlenecks bind.
• No effective slowdown. Amodei's "We Must Pace the Frontier" was endorsed by Altman and Musk, and labs' own policies call for halts at their critical thresholds.
• A risk report or update giving a point estimate of at least 100× with a method. Current reporting practice suggests this is plausible, but parts of methods are already redacted.

My estimates:

• An actual ≥100× speed-up at a lab inclined to publish, by about mid-2031: roughly 10–13%.
• Publishing a qualifying statement, given that speed-up: about 0.45.
• A generous survey-based or modelled claim without a real 100× speed-up: about 1%.

Together that gives about 6–7% by the end of 2031.

NO paths:

• Bottlenecks keep the speed-up at 3–30×.
• Coordinated slowdowns or a "speed limit" on self-improvement.
• A lab reaches the speed-up but stays vague or secretive, or only gives labour and code counts, which the question excludes.
• Publication lag runs past the deadline.

Shape of the series

The first quarter is nearly zero, since the measured speed-up is under 2× today. The quarterly hazard rises through 2028 and peaks in 2029, about six months after the aggressive forecasters' median takeoff to allow for publication lag. It then declines as slow-world scenarios dominate what remains. I built the series from 21 quarterly hazards between 0.10% and 0.48% and took the cumulative product.

Strongest case against my estimate

In this world things are moving at roughly 70–90% of the pace of the "AI 2027" scenario. Labs are unusually transparent, and the aggressive model gives about 50–60% to a 100× speed-up by 2031. If that is right and labs keep publishing, the true answer could be 15–20%. I weight it down for compute and software limits, explicit slowdown efforts, the strict method requirement, and labs' consistent caution about overall-pace claims (Anthropic said under 2× despite 4× self-reports and 8× code).

Sources

• OpenAI, 6 Sept 2026: https://openai.com/index/research-acceleration-view-inside-openai/
• Anthropic Risk Report, Aug 2026: https://www-cdn.anthropic.com/f61d49fa5596956a5dec75fea0e973bf6a6a8378/Redacted%20Risk%20Report%20August%202026%20.pdf
• Anthropic, 17 Sept 2026: https://www.anthropic.com/institute/measuring-pace-of-ai-development
• Anthropic, "When AI builds itself": https://www.anthropic.com/institute/recursive-self-improvement
• Gemini 4 safety report, Oct 2026: https://storage.googleapis.com/deepmind-media/gemini/gemini_4_safety_report.pdf
• Amodei, Sept 2026: https://darioamodei.com/post/we-must-pace-the-frontier
• CoinDesk, 12 Sept 2026: https://www.coindesk.com/tech/2026/09/12/anthropic-ceo-calls-for-ai-race-to-slow-down-musk-and-openai-s-altman-agrees
• Intelligence-explosion paper: https://casp.ac/reports/intelligence-explosion
• Forethought: https://www.forethought.org/research/how-quick-and-big-would-a-software-intelligence-explosion-be
• AI Futures, 16 Aug 2026: https://blog.aifutures.org/p/q25-2026-timelines-update-uplift
• METR, 14 Oct 2026: https://metr.org/blog/2026-10-14-better-informing-the-public-about-loss-of-control-risks/
• Kalshi API, event KXRSIANNOUNCE-29
• Import AI 476: https://importai.substack.com/p/import-ai-476-creativity-and-rsi

Question. By the last day of each quarter, has a covered developer stated, giving a number and a method, that AI systems have made the overall pace of its own AI research at least 100 times faster than it would be with human researchers alone?

Definitions.

• Covered developer: as defined by the site convention.
• Own AI research: the developer's internal research to develop or improve AI models and systems, including research engineering. AI systems used for this work may come from any developer.
• Overall pace: research progress per unit of elapsed calendar time across the developer's AI research program as a whole, not the speed of a selected task, team, experiment, or project. A sampled or modeled estimate counts only if the developer explicitly applies it to its overall program and describes how it aggregates or extrapolates to that scope. For a diversified parent company, the relevant scope is the covered AI developer, not all of the parent's research.
• Human researchers alone: a stated counterfactual in which humans perform the research without AI research assistance or autonomous AI researchers. Ordinary software, computation, and training the AI models being researched remain available. The method must identify the assumed human workforce and non-AI resources and distinguish the claimed AI contribution from unrelated changes in staffing, compute, or funding; identical total spending is not required.
• At least 100 times faster: an overall pace ratio of at least 100:1, or equivalent research progress achieved in at most one hundredth of the counterfactual time. A stated point estimate of at least 100 qualifies even with uncertainty bounds below 100. A range without a point estimate qualifies only if its lower endpoint is at least 100. “Up to 100 times” does not qualify.
• Giving a number and a method: the developer publishes the ratio, or numerical quantities directly yielding it, together with a description of the progress measure, actual research period, human-only baseline, and procedure used to estimate the comparison. Controlled comparisons, historical comparisons adjusted for other changes, researcher surveys, and counterfactual models can qualify. A bare assertion that an estimate was made, without describing how, cannot. Supporting methods may be in publicly accessible documents explicitly linked or cited by the statement; raw data and independent replication are not required.

Resolves YES if a qualifying developer statement reports an achieved acceleration meeting all definitions above. The statement must concern actual internal AI research, although the human-only comparison may be estimated. The developer must adopt the overall estimate as its own finding, rather than merely quote an outside forecast. The observation period convention applies to the research period underlying the estimate.

Resolves NO / Exclusions.

• Predictions, targets, hypothetical scenarios, or claims about what a system could achieve without a finding about actual research.
• Speedups solely in coding, experiment execution, inference, training throughput, or another component, unless accompanied by a qualifying overall research-pace estimate.
• Counts of AI researchers, human-equivalent labor, generated code, experiments, papers, or compute alone, without a method relating them to overall research progress and elapsed time.
• Comparisons against researchers already using AI assistance unless a separate qualifying human-only comparison is also given.
• Statements without a qualifying number or method, and statements only about another organization's research.
• No qualifying publication by the quarter-end.

Resolving source. The covered developer's own official publications only: statements on its domain or developer-authored reports, technical papers, system cards, or equivalent documents satisfying Public confirmation (A2). Third-party reports alone do not qualify. Apply the Resolution window and Observation period conventions. Resolve under Resolution date (A3) on the publication date when both the qualifying statement and its required method are public.