🔍 Read the full analysis: AI Improving AI: What Might The Process Change? on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
METR said on November 5 that Anthropic is likely improving model capabilities on well-specified tasks at about twice its earlier rate, with high uncertainty and AI assistance to researchers as the main contributor. Google DeepMind separately reported that AlphaEvolve made one matrix-multiplication component used in Gemini 23% faster, reducing overall training time by 1%. The findings suggest AI can help improve AI, but do not establish a self-sustaining or fully autonomous cycle.
An independent assessment published November 5, 2026, by the evaluation organization METR found that Anthropic is likely improving model capabilities on well-specified tasks at about twice its earlier rate, with high uncertainty. METR said the work it examined was primarily AI assistance to researchers, not a fully autonomous system improving itself. A separate Google DeepMind report said its AlphaEvolve system sped up one matrix-multiplication component used in Gemini, illustrating how AI-generated improvements could contribute to more efficient model development.
METR based its assessment on internal information from Anthropic covering August through October 2026. Its judgment was that the rate of capability improvement on well-specified tasks was around twice the earlier rate. METR stressed the uncertainty: it said it would not be surprised by estimates ranging from 1.3 times to three times. The assessment attributed most of the apparent acceleration to AI assistance in research, but the public account does not provide all the underlying evidence.
The source article says Anthropic redacted much of the information METR considered most informative. That limits outsiders’ ability to evaluate the assessment independently. The finding is an evaluator’s judgment about a particular organization and period, not a measurement that establishes the pace of AI research across the industry.
In a separate example, Google DeepMind reported that AlphaEvolve made a matrix-multiplication component 23% faster. The component is used in Gemini, and DeepMind said the improvement reduced Gemini’s overall training time by 1%. The distinction between the component-level gain and the end-to-end result matters: speeding up one part of a system does not produce an equally large reduction in the time or compute needed for the whole system.
How Research Gains Could Compound
The significance is not that these reports prove an imminent leap to superintelligence. It is that AI systems may help with parts of the process that produces more capable AI: reading research, writing or checking code, proposing experiments, and investigating results. If those contributions are reliable, researchers could test more ideas or spend less time on routine work, potentially increasing the output of a research team.
Efficiency gains can also add up, though examples should be treated as illustrations rather than forecasts. The source article gives a simplified calculation: if three independent improvements each save 10% of a task’s compute, the task would use about 73% of the original compute. Under the assumptions in that example, the same budget could support roughly 37% more such jobs. Actual gains depend on whether improvements are independent, apply to the same workload, and survive testing at larger scales.
The practical stakes extend beyond research labs. Faster or cheaper development could affect the timing of new AI capabilities and the investments organizations make in computing, staffing, and deployment. The distribution of any gains would also depend on who can use the systems, who owns the infrastructure, and how institutions share the benefits—questions that technical performance alone cannot answer.
Products Worth Considering
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black
Amazon Kindle Scribe (32GB) with Premium Pen, Tungsten (Like-New)
Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Better Models
Improving an AI system involves more than generating a promising idea. A researcher may need to review prior work, inspect an implementation, prepare data, run experiments, diagnose failures, and determine whether a result is real. AI can assist at several stages, but a reported improvement can be misleading: a test may have become easier, runs may use different amounts of compute, or a gain may disappear when the experiment is repeated with another random seed.
For that reason, useful research assistance includes testing proposals fairly and checking the evidence, not just producing code or a positive result. Partial automation could still matter if it lets researchers investigate more ideas or identify routine problems sooner. Greater autonomy would require systems to take on more of the sequence, including deciding what to investigate, while producing results that hold up to independent checks.
The METR assessment is a recent example of an attempt to evaluate whether AI research assistance is affecting the pace of development. The AlphaEvolve result offers a distinct, concrete example of an algorithmic change affecting a component of model training. Together, they point to possible links between AI tools and AI development, but they are different kinds of evidence and do not by themselves establish a general, closed feedback loop.
What the Evidence Cannot Yet Show
The public account does not include much of the evidence METR judged most informative because Anthropic redacted it. The assessment therefore cannot be fully checked by outside readers from the material described here. METR’s estimate also covers a particular period and organization; the source does not establish that similar acceleration is occurring across other AI laboratories.
It remains unclear how much of the reported progress came from AI assistance compared with other changes in research practice, and how reliably AI contributions hold up under repeated, independent testing. The AlphaEvolve example concerns one component and a reported 1% reduction in Gemini’s overall training time. It does not establish a continuing sequence in which each model independently builds a more capable successor.
Several constraints could slow the process: experiment time and available hardware, the difficulty of transferring results from small tests to large systems, choosing promising research directions, continuity across projects, and data quality. The source’s broader claim that progress could become rapid is a possibility, not a confirmed timetable. The pace could vary rather than follow a smooth curve.
Testing Gains at Larger Scale
The next useful evidence would show whether AI-assisted research gains persist across more tasks, research teams, and time periods, with enough disclosed methods for outside evaluators to assess them. Repeated tests should distinguish genuine improvements from changes in experimental setup and report how much human supervision remains necessary.
For efficiency claims, readers will need both the local result and the effect on the full system, as the reported 23% component improvement and 1% training-time reduction illustrate. Further reporting could clarify whether such gains recur and whether they reduce costs or increase research output in practice.
For now, the reports support a measured conclusion: AI is contributing to some research and engineering work involved in developing AI, and those contributions may affect the pace of progress. The scale, durability, and broader consequences of that effect remain open questions.
Key Questions
What did METR conclude about Anthropic’s AI research progress?
METR judged that Anthropic’s capabilities on well-specified tasks were improving at around twice the earlier rate. It described the estimate as highly uncertain, saying it would not be surprised by values from 1.3x to 3x.
Did METR find that an AI system was improving itself autonomously?
No. The source says the work METR described was primarily AI assistance to researchers. The assessment does not establish a fully autonomous system designing and building its own successor.
What did AlphaEvolve change in Gemini?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing the model’s overall training time by 1%.
Why is the METR assessment difficult to verify independently?
According to the source article, Anthropic redacted much of the evidence METR considered most informative. That means external readers have limited access to the material behind the evaluator’s judgment.
Does this evidence show how quickly AI will advance?
No. The findings concern a specific organization, period, and engineering result. They suggest that AI assistance may affect research pace, but the scale, repeatability, and future trajectory remain uncertain.
Source: ThorstenMeyerAI.com
