🔍 Read the full analysis: AI’s Role In Building The AI That Comes Next 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
An independent METR assessment published Nov. 5 says Anthropic is likely improving model capabilities on well-specified tasks at about twice its earlier rate, with high uncertainty. METR judged AI assistance to researchers was the main contributor to the acceleration it assessed; a separate Google DeepMind example shows an AI-designed change sped up part of Gemini training. How far these gains can scale remains unclear.
AI assistance may be speeding up some model-development work, according to an assessment of Anthropic published Nov. 5 by the independent evaluation organization METR. METR judged that the company was likely improving model capabilities on well-specified tasks at about twice its earlier rate, while emphasizing substantial uncertainty; the finding matters because AI systems may help researchers build the models that follow them.
METR’s assessment drew on internal information from August through October 2026. It said it would not be surprised if the rate were as low as 1.3 times or as high as three times the earlier rate. METR judged that AI assistance to researchers was likely responsible for most of the acceleration it observed. The account describes assistance, not a fully autonomous system independently conducting AI research.
The public record has limits. Anthropic redacted much of the evidence METR considered most informative, so readers cannot inspect all the material behind the assessment. METR’s conclusion is an evaluation based on evidence it reviewed, not a public, independently reproducible demonstration of a specific amount of acceleration.
A separate example cited by Google DeepMind concerns AlphaEvolve, which the company said produced a matrix-multiplication component used in Gemini that ran 23% faster. DeepMind reported that this change reduced Gemini’s overall training time by 1%. That distinction matters: a substantial improvement to one component did not translate into an equally large reduction across the full training process.
How Research Gains Could Compound
The development matters because AI research contains many tasks that can be assisted by software: reading technical work, inspecting code, proposing experiments, analyzing results and finding implementation errors. If a system can do some of those tasks reliably, researchers may be able to test more ideas or spend less time on routine work. That could raise the pace of development even without a model taking over an entire research program.
Efficiency improvements can also add up. As a simplified illustration, three independent savings of 10% would leave a job requiring about 73% of its original compute; on the same budget, that would allow roughly 37% more such jobs. This is an arithmetic example, not a measured result from either METR or DeepMind. Real projects may not have independent gains, and compute savings do not by themselves establish faster scientific progress.
For companies, workers and policymakers, the practical issue is that planning assumptions could change if AI helps shorten parts of the research cycle. But technical capability alone does not determine who can use these systems, who controls the computing infrastructure or how resulting benefits are distributed. Those outcomes also depend on deployment choices and institutions.
Products Worth Considering
Samsung Odyssey G7 G75F 40-Inch WUHD Curved Gaming Monitor
DJI Osmo 360 Motorcycle Handlebar POV Combo
Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Successors
The relevant feedback loop is straightforward: researchers use an AI model to help with development; some suggestions pass testing and are incorporated; a later model may then assist with subsequent work. The tool being improved could contribute to its own improvement, but this does not mean the process is automatic or guaranteed to accelerate indefinitely.
Research work involves more than producing a promising code change. Teams must compare experiments fairly, check that results hold across conditions and distinguish genuine progress from errors. A reported improvement could reflect a changed test, unequal computing resources or random variation. Independent checks and careful measurement are therefore part of the research task, not optional extras.
The source material also points to possible constraints: experiment duration and hardware availability, the difficulty of scaling results from small trials to large systems, selecting useful research directions, maintaining continuity and obtaining good data. These factors could limit or interrupt gains, even where AI assistance is useful.
“METR said it “would not be surprised” by an acceleration of 1.3x or 3x.”
— METR, in its assessment of Anthropic published Nov. 5, 2026
Limits in the Public Evidence
The estimate for Anthropic is explicitly uncertain, and the redacted evidence prevents outsiders from reviewing much of what METR found most informative. The assessment concerns improvements on well-specified tasks; the source does not establish that all AI research tasks are accelerating at the same rate.
It is also not clear how much of the reported pattern would persist over longer periods, how much could be reproduced at other laboratories, or whether current assistance will extend to independently choosing research directions and validating results. The source material provides no common measure that directly compares the METR assessment with the AlphaEvolve example. Claims about rapid progress toward superintelligence remain a possibility raised by the author, not a conclusion established by these examples.
What Future Assessments Must Show
The next useful evidence would include clearer descriptions of what tasks AI systems perform, how much human supervision they require and whether gains survive independent testing. Further assessments over time could help determine whether the reported acceleration is sustained, limited to certain work or affected by changing methods and resources.
For now, the confirmed picture is narrower: METR judged that AI-assisted work likely contributed to faster capability improvement on specified tasks at Anthropic, with high uncertainty, while DeepMind reported a specific efficiency gain in part of Gemini’s training process. How far those results scale—and whether they change the overall pace of AI development—remains an open question.
Key Questions
What did METR conclude about Anthropic?
METR judged that Anthropic was likely improving model capabilities on well-specified tasks at about twice its earlier rate. It emphasized high uncertainty, giving a range of 1.3x to 3x as possibilities it would not be surprised by.
Did METR find that AI systems are conducting research autonomously?
No. The source describes the work METR assessed as primarily AI assistance to researchers. It does not establish that a system independently ran an entire research program.
What did AlphaEvolve change for Gemini?
Google DeepMind said AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster. DeepMind reported that the change reduced overall Gemini training time by 1%.
Why is the Anthropic assessment difficult to verify publicly?
Anthropic redacted much of the evidence METR considered most informative. The assessment’s conclusion is public, but outsiders cannot review all of the underlying material described in the source.
Does this evidence establish that AI development will keep accelerating?
No. The assessment concerns a specific company and category of tasks, and METR described substantial uncertainty. Hardware, experiment time, research choices and the challenge of scaling results could all affect future progress.
Source: ThorstenMeyerAI.com
