takeoff.watch

An independent evaluator finds AI has made a lab's research 100× faster

Will an independent organisation publish an estimate, giving a number and a method, that AI systems have made the overall pace of a covered developer's AI research at least 100 times faster than it would be with human researchers alone?

Chance by Q4 2031

9%

Cumulative probability

0%25%50%75%100%
  • Latest forecast

Forecast history

Forecast made on 14 Oct 2026

Latest of 2 forecasts · Methodology v2.1

Model reasoning

Forecast made 14 Oct 2026 · Methodology v2.1

The best independent estimate so far says artificial intelligence makes a leading lab's research about one and a half times faster than its staff alone. A hundredfold speedup within five years would take something like an intelligence explosion, and an outside evaluator would then have to publish a careful figure. So this is unlikely but possible. The biggest change would be evidence that machines are running research largely on their own, with the measured speedup climbing fast.

Where things stand (mid-October 2026)

  • The only independent estimate is about 1.5x. METR's summary of its pre-deployment evaluation of Claude Opus 5.5 (22 Sep 2026) quotes a preliminary report from a separate METR team with elevated access: "~1.5X overall acceleration in capabilities due to AI (i.e. 1.5 years in 1 year), with perhaps 30% chance of 2X." It gives no time period and no public method, so it would not qualify here even at a higher number. METR says more public outputs from that investigation are coming "in the coming weeks."
  • Anthropic's own numbers are also below 2x. Its Responsible Scaling Policy v3.4 threshold is a doubling of aggregate capability progress. The Opus 5.5 system card finds the slope has not doubled: a trend-break reading gives 1.53x (95% range 1.20–1.82), and a one-time-jump model fits better. Earlier, the Mythos system card said acceleration was "well short of" 2x. In METR's Frontier Risk Report (May 2026), Anthropic said a 4x increase in labour input maps to only about 1.7x more output.
  • Other estimates are similar. A July METR note by Thomas Kwa puts researcher uplift at about 2.5x, which implies roughly 1.8x overall research speedup. The Elasticity Institute paper (July 2026) estimates the self-improvement feedback at about 9%, below the roughly 15% needed for self-sustaining acceleration.
  • So the threshold is about 67 times today's level. 100x is roughly AI 2027's "superhuman to superintelligent AI researcher" range.

Reference class and model

There is no direct base rate; no lab has ever been credibly estimated above about 2x. I treat YES as the product of two things.

1. The real multiplier reaches at least 100x at a covered developer.

  • AI Futures Project. I downloaded their model's all-things-considered forecasts (aifuturesmodel.com, Aug 2026 versions). In Daniel Kokotajlo's median trajectory the software progress multiplier passes 100x around 2028.6, at about the "top-expert-dominating AI" milestone (about 101x). Their probabilities that this milestone arrives by the start of 2032 are 84% (Daniel), 48% (Eli) and 55% (Brendan). These are explicitly conditional on no deliberate slowdown, and this group sits at the aggressive end.
  • Metaculus. Weakly general AI is forecast around Aug 2027, and the stronger general-AI question around Jan 2031 (interquartile range 2028–2037).
  • Skeptics. Ramez Naam (Sep 2026), Arvind Narayanan, and analysts who stress compute bottlenecks would put 100x by 2031 at a few percent.
  • Pacing. Dario Amodei's "We Must Pace the Frontier" (12 Sep 2026) proposes a "speed limit" on recursive self-improvement, and Altman has backed parts of it. This cuts the chance that an observed 100x happens at US labs.
  • My blended curve for a real 100x: about 2% by end-2027, 7% by end-2028, 11% by end-2029, 15% by end-2030 and 18% by end-2031.

2. A qualifying publication follows. It needs an independent organisation with an access agreement, a central estimate of at least 100, a named developer, a published method, and an observed (not projected) figure.

  • What helps: METR is funded (about $71M) to track recursive self-improvement. Anthropic, with OpenAI saying it will match, is committing to embedded evaluators who have the right to publish findings.
  • What hurts: government secrecy, labs without evaluator arrangements, hedged ranges whose lower bound falls below 100, and the lag between the speedup happening and a report appearing.
  • My numbers: about 49% eventual qualifying publication, with a lag of a few quarters (8% in the same quarter, 40% within four quarters). I add a small hazard (about 1% over five years) for a qualifying estimate of at least 100 when the real figure is lower.

Computation. I took the per-quarter chance of first reaching a real 100x from the curve above, applied the publication-lag weights quarter by quarter, and combined the result with the small extra hazard (Python). Cumulative chance of YES:

By end of 2027 2028 2029 2030 2031
Chance of YES 0.5% 2.2% 4.5% 6.7% 8.6%

The per-quarter hazard peaks around 0.6% in 2029–2030.

How I read the question. "Overall pace" means a developer-wide research-progress multiplier with compute and staff held fixed, in the same sense as Anthropic's policy threshold and AI 2027's "progress multiplier."

Strongest case against my estimate

Lab insiders sound much more bullish than this. Amodei says AI has been "advancing drastically faster" since the summer. OpenAI's chief scientist has a "strong expectation" of recursive self-improvement, and OpenAI targets an automated AI researcher by March 2028. Kokotajlo reports that insiders call AI 2027 too conservative. If takeoff starts in 2027–28, embedded evaluators could report 100x or more by 2029, which would put this near 15–20%.

I don't go that high, for three reasons:

  • Measured data still show about 1.5x with no clear acceleration of the acceleration.
  • Research judgement remains weak in recent METR and Epoch evaluations.
  • Compute bottlenecks and pacing push against an observed 100x, and the publication bar is strict.

Sources

Resolution criteria

Opened 7 Oct 2026Not yet resolved

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

Definitions.

  • Independent organisation: a named organisation that is not part of, controlling, controlled by, or under common control with the covered developer concerned (control meaning a majority of equity, votes, or board appointments), and neither it nor any entity controlling it has made a direct equity investment in that developer. A government body is assessed apart from its government's other holdings. Paid work for the developer, compute or model access, grants, donations, and the developer's pre-publication review or redaction do not affect independence; an author employed by the developer does. It must have produced the estimate with access to the developer's non-public research data, staff interviews or surveys, or unreleased models under an agreement or memorandum of understanding with the developer or a legal or regulatory requirement, as stated in a publication by either.
  • Overall pace of a covered developer's AI research: the developer-wide rate of substantive progress in developing or improving AI models and methods, measured per unit of calendar time. For a covered developer that is a named AI lab or model organization within a larger company, the scope is that organization's AI research, not the parent's unrelated work. An aggregate estimate from a sample counts only if the organisation explicitly presents it as an estimate of the developer's overall AI research pace.
  • Human researchers alone: a counterfactual in which the same human research workforce has ordinary non-AI software and research infrastructure but no AI systems performing or assisting research work. This does not exclude training or evaluating AI models as objects of research. The comparison must distinguish AI assistance from changes in human staffing, training compute, or other non-AI resources; a historical baseline counts only if the method explains how those differences are accounted for.
  • At least 100 times faster: a ratio of AI-assisted research progress per unit time to the human-only counterfactual of at least 100.0, or equivalent time to the same progress of at most one-hundredth. The AI systems need not have been built by the developer whose research is estimated.
  • Giving a number and a method: the publication reports the multiplier (or quantities explicitly equated to it), identifies the research-progress measure and observation period, and explains how both the AI-assisted result and human-only counterfactual were estimated and aggregated to the developer-wide level. Experimental comparisons, observational analyses, and explicitly described estimation models may count; public raw data, independent replication, and statistical significance are not required. Survey evidence counts only when the organisation uses a disclosed method to translate it into and endorse an overall research-progress estimate, rather than merely reporting respondents' output estimates.

Resolves YES if a resolving source gives, as the independent organisation's own estimate, an achieved or currently observed overall multiplier of at least 100.0, explicitly applied to a named covered developer, with the method published there or in public documentation it cites. A central or point estimate counts despite an uncertainty interval extending below 100.0; “about,” “roughly,” or “~” figures are read at face value; a range given alone is read at its lower bound; and if several estimates for the same developer and period include none identified as central or overall, the lowest governs. Preliminary or experimental estimates count.

Exclusions (resolve NO).

  • Multipliers for coding speed or volume, experiment counts, or agent labor (such as agent-workdays per human workday).
  • Unadjusted averages or medians of staff's self-reported output gains.
  • Figures such as “up to 100x” or “nearly 100x,” and a stated probability that the multiplier reaches 100.

Resolving source. Only the organisation's website, including staff-authored notes hosted there; reports, papers, preprints, testimony, or submissions under its name or affiliation; and sections of developer-published documents explicitly presented as its own text, not a developer's paraphrase, summary, or quotation. This replaces Public confirmation (A2): developer disagreement does not prevent YES, but the organisation withdrawing the estimate, or revising it below 100.0, within 14 days of publication does. Apply the Resolution window, Observation period, and Resolution date (A3) conventions; the date is when the estimate and its method are both public.

Download all forecasts (JSON)