🔍 Read the full analysis: When AI Becomes A Tool For Improving AI on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
METR’s November 5 assessment judged that Anthropic’s AI capabilities on well-specified tasks were improving at about twice the earlier rate, with substantial uncertainty. It attributed most of the acceleration it examined to AI assistance for researchers, while noting that much of the underlying evidence was redacted. A separate Google DeepMind example shows how a targeted efficiency improvement can reduce training time, but does not establish that AI research can accelerate without limit.
Independent evaluator METR said in a November 5 assessment that Anthropic was likely improving its AI capabilities on well-specified tasks at about twice its earlier rate, with substantial uncertainty. METR judged that AI assistance to researchers was likely responsible for most of the acceleration it examined, putting a concrete but qualified estimate on the possibility that AI tools can help develop more capable AI systems.
METR’s assessment drew on internal information from August through October 2026. The organization said it would not be surprised if the rate were as low as 1.3 times the earlier pace or as high as three times. Its finding concerns capability gains on well-specified tasks; it is not a claim that every part of AI research, or the whole development process, has sped up by the same amount.
The source article says the work METR described was primarily AI assistance to human researchers, rather than fully autonomous research. In practical terms, systems may help with tasks such as reading technical work, editing code, setting up experiments, investigating failed runs, and checking whether a result holds up. The source emphasizes that a useful contribution must survive fair testing: apparent gains can result from mismatched compute, easier tests, or random variation.
There is a significant limit to what outsiders can independently check. According to the source article, Anthropic redacted much of the evidence METR considered most informative. The public can therefore see the evaluator’s conclusion and stated uncertainty, but not all the details behind it. The assessment is evidence of reported acceleration, not a complete public record of the methods or measurements.
How Research Gains Could Compound
If AI systems help researchers test more ideas or spend less time on routine coding and debugging, a team may be able to make progress faster without automating every research decision. If better models then contribute more effectively to the next round of work, improvements could reinforce one another. That feedback loop is the central reason the source argues that AI-assisted research deserves attention.
A separate example cited by the source comes from Google DeepMind: it reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing Gemini’s overall training time by 1%. The figures refer to a particular component and its effect on overall training, not a 23% reduction in the full training run. They illustrate how a substantial local improvement can translate into a smaller system-wide saving.
The source also offers a simplified illustration, not a measured result: three independent 10% reductions in resource use would leave a job needing about 73% of its original compute. Under that set of assumptions, the same compute budget could support roughly 37% more such jobs. This shows why efficiency gains may matter for research capacity, but it does not establish that real-world gains are independent or will accumulate at that rate.
Popular right now
Vitamix Touch & Go Advance 32-Ounce Bar Blender, White
De'Longhi Stilosa Compact Manual Espresso Machine with Milk Frother
De'Longhi Classic Signature Espresso Machine
As an affiliate, we earn on qualifying purchases.
From AI Assistance to Research Acceleration
The development described is a shift in how AI may be used: not only as the subject of research, but also as a tool in the work of improving AI. Researchers can give systems access to code, experiment results, and a defined problem, then use their help to propose or test changes. Human oversight remains relevant, particularly when deciding whether an apparent improvement is genuine and worth adopting.
METR’s assessment provides a recent estimate of this activity at Anthropic, while the AlphaEvolve example concerns an efficiency change reported by Google DeepMind. They are distinct examples and should not be treated as a single study or as proof of a general trend across laboratories. Together, they show why researchers and evaluators are examining both the pace of capability development and the resources needed to achieve it.
The source identifies several possible limits on further acceleration: the time and hardware required to run experiments; whether results from small tests transfer to large systems; how teams choose research directions; continuity in the work; and data quality. Automating individual tasks does not by itself resolve those constraints or establish that a system can direct a research program reliably.
“It “would not be surprised” by a rate of 1.3x or 3x.”
— METR, as described in its November 5, 2026 assessment
What the Public Evidence Cannot Show
The estimate remains qualified by substantial uncertainty, and the source says METR’s most informative evidence was partly redacted by Anthropic. The public material described here does not supply the full underlying data, allowing readers to independently reproduce the assessment, or explain every step by which the evaluator reached its judgment.
It is also unclear how broadly the estimate applies beyond the specific tasks and period METR examined. The assessment does not establish that AI assistance caused all capability gains, that systems can independently choose valuable research directions, or that the pace will continue. The source’s efficiency examples likewise do not show that gains will reliably compound across different models, experiments, or laboratories.
More broadly, technical progress alone does not determine who can use these systems, who controls the computing infrastructure, or how resulting benefits are shared. Those outcomes depend on deployment and institutional choices as well as model capabilities.
Evidence to Watch in Future Assessments
The next useful milestone will be further public evidence about how AI assistance changes research output: which tasks systems perform, how much human supervision they need, and whether claimed improvements hold under independent checks. More transparent methods and results would make it easier to judge how much weight to place on METR’s estimate.
Researchers and decision-makers will also need to track whether improvements transfer from small experiments to large training runs, and whether bottlenecks such as hardware, experiment time, and research direction limit the effect. For now, the reported assessment supports a cautious conclusion: AI assistance may be speeding parts of AI development, but its scale, durability, and wider consequences remain unsettled.
Key Questions
What did METR report about Anthropic?
METR judged that Anthropic’s capabilities on well-specified tasks were improving at about twice the earlier rate. It gave a wide range of plausible rates, from 1.3 to three times, and said the estimate carried substantial uncertainty.
Did METR say AI is conducting research autonomously?
No. The source article says the work METR described was primarily AI assistance to researchers. The assessment does not establish that AI systems are independently directing a full research program.
What did AlphaEvolve change in Gemini?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall training time by 1%. The component-level figure is not the same as a 23% reduction in total training time.
Why is the METR estimate difficult to verify publicly?
According to the source article, Anthropic redacted much of the evidence METR considered most informative. That means outside readers do not have access to all the underlying details needed to independently assess the estimate.
Does the report show that AI progress will keep accelerating?
No. It describes an assessment of a particular period and set of tasks, with substantial uncertainty. Experiment time, hardware, data quality, and the challenge of transferring small-scale results to larger systems may all affect future progress.
Source: ThorstenMeyerAI.com
