A lab says AI has made its research 20× faster

Where things stand (as of 16 Oct 2026)

Nothing published since the 16 Sept 2026 opening date qualifies. Every developer that has put a figure on its overall pace reports less than 2×:

• Anthropic, Risk Report Aug 2026, §3.5.2. "Meaningful acceleration… by less than a factor of 2." Anthropic says it has deprioritized its staff surveys (geometric-mean self-reported uplift about 4×) because "productivity uplift on individual tasks does not translate directly into acceleration of research progress." Its policy threshold for "dramatic acceleration" is only 2×.
• Claude Opus 5.5 System Card, 22 Sept 2026. "We do not observe a sustained AI-attributable 2× acceleration." Its capability-index slope rose from 14.4 to 22.2 points per year, about 1.5×. The card includes METR's preliminary estimate of "~1.5X overall acceleration in capabilities due to AI (1.5 years in 1 year), with perhaps 30% chance of 2X."
• Google DeepMind, Gemini 4 safety report, Oct 2026, §3.4.2. "No significant end-to-end acceleration." Its capability-index slope went from 0.091 to 0.118 points per month, which is not statistically significant.
• OpenAI, "Research acceleration: the view inside OpenAI," 6 Sept 2026. Reports input metrics only (3.1 agent-workdays per human workday, experiments per researcher) and gives no overall multiplier. It warns that "the overall pace of progress likely won't keep pace with these specific metrics."
• Anthropic Institute pages, 17 Sept 2026 (pace measurements) and June 2026 ("When AI builds itself"). These report automation shares (Claude "leads" 26% of tasks; above 90% at "collaborates" or higher), about 30,000 agents, and 8× code per engineer. All of these are excluded categories under the resolution criteria.
• METR, 14 Oct 2026. "No company has released detailed public evidence" on overall acceleration.

Reading adopted

A qualifying statement needs an organization-wide multiplier on research progress or comparable output, measured against a human-only counterfactual with a stated method. One example would be a capability-trend acceleration, as in the risk reports, with AI's share separated from other causes. Survey uplift, code volume, automation shares and agent counts do not count unless they are explicitly translated into such a multiplier.

Reference class and model

Labs and their evaluators define acceleration as years of progress per calendar year, so 20× means about 20 years of progress per year. That is roughly the "superhuman AI researcher" stage in the AI 2027 scenario (about 25×), i.e. a full intelligence explosion.

Two trajectories bracket the timing:

• Trend extrapolation. If the excess over 1× (0.5 in mid-2026) doubles every 6–12 months, 20× arrives around 2029–2032.
• Forecaster views. AI Futures (16 Aug 2026) judges reality to be running at 70–90% of AI 2027's pace. Jack Clark (Anthropic) now favors 2027 for full self-improvement. On the other side, compute bottlenecks (OpenAI's "constant compute crunch"), skeptics such as Naam and Narayanan, the voluntary "pacing the frontier" push, and OpenAI's own Critical threshold (5×), which calls for halting development, all pull later.

How the forecast was built (Python, in the scratch container; inputs are my judgments informed by the sources above):

1. The chance that true acceleration of 20× or more first occurs at some covered developer: about 3% by end-2027, 10% by end-2028, 20% by end-2029, 31% by end-2030 and 39% by end-2031.
2. Multiplied by a 0.5 chance that the developer then publishes a central estimate of at least 20× with a method. This allows for conservative measurement, redaction of leading indicators, possible national-security secrecy, and labs reporting against thresholds rather than giving point estimates.
3. A publication lag of 25% in the same quarter, 30% one quarter later, 20%, then 12%, then a decaying tail.
4. A small separate chance of a generous claim that still qualifies (0.06–0.15% per quarter).

The resulting quarterly risk of resolution is about 0.1–0.4% in 2027, rising to about 1.5% per quarter in 2030–31.

Strongest counterarguments

• For higher. Insiders (Amodei, Pachocki, Clark) expect self-improving AI soon. Anthropic has a policy incentive to publicize acceleration to support pacing. Transparency norms are strengthening (embedded evaluators, the White House accord, OpenAI's call for standards on measuring self-improvement progress). If self-improvement arrives in 2027–28, a 20× reading could follow by 2029 and still be published inside the window.
• For lower. Twenty-fold capability progress is extreme, and capability indices saturate. Compute is held fixed in the counterfactual. Labs measure conservatively: Google still sees no acceleration at all.

I weighed both and landed at about 20% by end-2031.

Sources

• https://www.anthropic.com/institute/measuring-pace-of-ai-development
• https://www.anthropic.com/institute/recursive-self-improvement
• https://www-cdn.anthropic.com/f61d49fa5596956a5dec75fea0e973bf6a6a8378/Redacted%20Risk%20Report%20August%202026%20.pdf
• https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf
• https://metr.org/blog/2026-09-22-claude-opus-5-5/
• https://storage.googleapis.com/deepmind-media/gemini/gemini_4_safety_report.pdf
• https://openai.com/index/research-acceleration-view-inside-openai/
• https://openai.com/index/building-standards-next-phase-ai/
• https://metr.org/blog/2026-10-14-better-informing-the-public-about-loss-of-control-risks/
• https://darioamodei.com/post/we-must-pace-the-frontier
• https://blog.aifutures.org/p/q25-2026-timelines-update-uplift
• https://www.therundown.ai/news/jack-clark-ai-self-improvement-forecast-2027

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 20 times faster than it would be with human researchers alone?

Definitions.

• Covered developer: as defined by the site convention. For a developer within a larger corporate group, the relevant scope is that developer's AI research organization, not the parent's unrelated businesses.
• Own AI research: research and research engineering directed toward developing or improving AI models, including designing, implementing, running, and interpreting experiments. AI systems used for this work need not have been developed by the reporting developer.
• Overall pace: the rate of research progress or comparable research output per unit of elapsed time across the developer's AI research organization. The developer must explicitly characterize the estimate as organization-wide or as an estimate of its AI research pace overall. A representative sample or aggregate model can support that estimate; a speedup limited to one task, project, team, or research specialty cannot qualify without an explicit method of extrapolation to the overall pace.
• Human researchers alone: a counterfactual in which the developer's human research workforce works without AI research assistance, but retains conventional software, experimental equipment, and model-training compute. The comparison must isolate the contribution of AI assistance rather than simply compare the present organization with a smaller or less-resourced historical organization. A historical baseline counts if the method explains how other changes are controlled for or separated from AI's contribution.
• At least 20 times faster: an AI-assisted research rate at least 20 times the human-only rate, equivalently no more than one-twentieth of the time for comparable research progress. An explicitly reported central estimate of 20× or more counts even if its uncertainty interval extends below 20×. A range without a central estimate counts only if its lower endpoint is at least 20×; an upper bound, possible maximum, or “up to 20×” does not count.
• Giving a method: the publication, or publicly available material it explicitly references, identifies the research-output or progress measure, the work and period assessed, how the human-only counterfactual was estimated, and how measurements or judgments were combined into the overall multiplier. Controlled comparisons, retrospective estimates, surveys, and modeled counterfactuals are eligible. Raw data, independent replication, and a particular statistical significance level are not required.

Resolves YES if a covered developer publishes an affirmative statement meeting all these conditions about acceleration already achieved in its own research. The method may estimate a counterfactual, but the AI-assisted work must be actual work, not merely a projected future deployment. Apply the Resolution window and Observation period conventions.

Resolves NO if no qualifying statement has been published by the quarter-end. Exclude benchmark performance, code-volume increases, numbers of agents, automated-work shares, human-equivalent labor estimates, and individual productivity figures unless the developer explicitly translates them into an overall research-pace estimate using the qualifying method. Exclude forecasts, capability claims that a model could accelerate research, and unsupported multipliers without a method.

Resolving source. Developer-authored publications only, under Public confirmation (A2): official research posts, technical reports, system cards, risk reports, or other developer-authored documents. Relevant publication channels include Anthropic's Institute and OpenAI's research publications; these are examples, not exclusive sources. A third-party estimate alone does not qualify. Apply Resolution date (A3); if the method is published later than the claim, use the publication date on which the complete qualifying evidence first becomes available.