🔍 Read the full analysis: When AI Starts Helping With Its Own Successor on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A November 5 METR assessment judged that Anthropic’s model capabilities on well-specified tasks were improving at about twice the earlier rate, with high uncertainty, and said AI assistance to researchers likely accounted for most of the acceleration. The report offers evidence that AI may help improve future AI systems, but it does not establish that research is fully automated or that progress will keep accelerating.
An independent assessment published by METR on November 5 judged that Anthropic was likely improving its models’ capabilities on well-specified tasks at about twice the earlier rate, with substantial uncertainty. METR said AI assistance to researchers likely accounted for most of the acceleration, offering a concrete example of AI systems helping with the research that could produce their successors.
METR said its assessment drew on internal information from August through October 2026. Its estimate was not presented as a precise measurement: the organization said it would not be surprised by a rate as low as 1.3 times the earlier pace or as high as three times. The work described was primarily AI assistance to researchers, not a finding that models were independently conducting the full research and development process, as explored in efforts to build AI that refines its own work.
There is a significant limit on what outsiders can verify. According to the source article, Anthropic redacted much of the evidence that METR considered most informative. The assessment therefore offers an evaluator’s judgment based partly on material the public cannot inspect. It indicates a possible change in the pace of work, but does not settle how much of that change was caused by AI or how widely the result applies.
A separate example cited in the source comes from Google DeepMind: the company reported that its AlphaEvolve system made a matrix-multiplication component used in Gemini 23% faster, reducing Gemini’s overall training time by 1%. That is a company-reported engineering result, not evidence by itself of a general acceleration in AI research. It does illustrate how a targeted improvement to a component can reduce the resources needed for a larger system.
How Research Gains Could Compound
If AI tools help researchers write code, test ideas, diagnose failed runs or check results, a team may be able to examine more possibilities with the same staff and computing budget. If some tested improvements then feed into a stronger model, that model could provide more effective assistance on later work. This creates a potential feedback loop: AI contributes to research, research improves models, and improved models may contribute further.
The size and reliability of each contribution matter. An apparent improvement can come from a flawed comparison, a change in test difficulty or random variation rather than a genuine advance. A useful system must do more than propose changes: researchers need to reproduce results, check that tests are fair and decide whether evidence justifies adopting a method. The METR assessment points to assistance as a factor in faster work, but does not show that these checks or human decisions have become unnecessary.
Even modest efficiency improvements could matter when repeated across expensive training work. The source article offers a simplified illustration: three independent 10% reductions in resource use would leave a job requiring about 73% of its original compute, allowing roughly 37% more jobs on the same budget. That is an illustration based on assumed independent savings, not a reported result from Anthropic or Google DeepMind. Whether savings can be repeated and combined in real systems remains an empirical question.
The broader stakes extend beyond model development. A sustained change in how quickly AI capabilities improve could affect how companies, workers and public institutions plan for new tools. But technical progress alone does not determine who can use those tools, who controls the infrastructure or how economic gains are shared; those outcomes also depend on deployment choices and institutions.
Products Worth Considering
Samsung 57 Odyssey Neo G9 (G95NC) Dual 4K UHD 1000R Curved Gaming Monitor
Samsung Odyssey G55C 32-Inch QHD 1000R Curved Gaming Monitor
LG UltraGear 45GX950A-B 45-Inch 5K2K Curved OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
What the Report Actually Measured
The phrase “AI helping build its successor” can imply a system independently choosing research goals, running experiments and producing a new model. The evidence described here is narrower. METR assessed changes in capability on well-specified tasks at Anthropic and judged that AI assistance to researchers likely explained most of the acceleration. The source does not describe a fully autonomous research process.
Research involves many separate steps: reviewing prior work, inspecting code, preparing data, setting up experiments, measuring results and investigating failures. AI can assist with some steps without being reliable across the whole sequence. For example, a code change that appears to improve performance may have made the test easier or used a different amount of compute. Repeating experiments and checking results can reveal those problems.
The DeepMind example concerns a different kind of evidence: a reported optimization to a component in Gemini’s training process and the resulting reduction in overall training time. Together, the two examples show distinct ways AI can support technical work, but they should not be treated as a single controlled measure of self-improving AI. METR’s assessment and DeepMind’s reported optimization have different methods, scopes and evidentiary limits.
“It would not be surprised by a rate of 1.3 times or three times the earlier pace.”
— METR
Evidence and Limits Still Matter
The public cannot fully evaluate METR’s assessment because, according to the source article, Anthropic redacted much of the evidence the evaluators found most informative. The source does not provide the underlying data, a detailed breakdown of which research tasks AI handled, or a publicly inspectable method for attributing the change in pace to AI assistance.
The estimate also carries a wide uncertainty range. It concerns capability progress on well-specified tasks, not every measure of AI performance or the overall pace of scientific discovery. It remains unclear how much the assessment generalizes beyond the period studied, whether the reported pace will persist, and how much of any gains came from AI rather than other changes in research practice or resources.
More broadly, no evidence presented here establishes that AI systems can reliably select the most valuable research directions, validate their own findings without human checks or independently build a successor model. Constraints may include experiment time, access to hardware, the difficulty of scaling small tests to large systems, continuity across work and the quality of available data.
What Future Assessments Need to Show
The next useful evidence would clarify what tasks AI systems perform in research, how much supervision they need and whether their contributions survive independent replication. Public details about evaluation methods and comparison baselines would help readers distinguish a genuine change in research productivity from a promising but narrow result.
Researchers and evaluators will also need to track whether gains persist as experiments move from small tests to large training runs, where time and hardware can limit how quickly ideas are checked. The METR assessment, as described in the source, concerns internal information from August through October and does not establish what happened after that period.
For now, the supported conclusion is limited: METR judged that AI assistance likely contributed to faster capability improvement at Anthropic, but with high uncertainty and evidence partly withheld from public view. Whether that develops into a durable, compounding cycle of AI-assisted research remains to be tested through further assessments and disclosed results.
Key Questions
What did METR conclude about Anthropic’s AI research?
METR judged that Anthropic’s models were improving on well-specified tasks at about twice the earlier rate and that AI assistance to researchers likely explained most of the acceleration. It attached substantial uncertainty to that estimate.
Did METR find that AI is building its own successor?
No. The assessment described in the source concerns AI assistance to researchers. It does not establish that a model independently directs research or builds a successor without human involvement.
Why is the assessment difficult for the public to check?
The source article says Anthropic redacted much of the evidence METR considered most informative. That leaves outsiders unable to inspect important supporting material and reliant on METR’s assessment.
What was the AlphaEvolve result?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, cutting overall training time by 1%. This is a company-reported engineering result, separate from METR’s assessment of research acceleration at Anthropic.
Does the reported acceleration mean AI progress will keep speeding up?
That is not established. METR reported high uncertainty, and constraints such as experiment time, hardware, research direction and the difficulty of scaling tests may limit further gains. More evidence is needed to determine whether the pace persists.
Source: ThorstenMeyerAI.com
