TL;DR
METR’s November 5 assessment said Anthropic is likely improving model capabilities on well-specified tasks at about twice its earlier rate, with high uncertainty; it attributed the work primarily to AI assistance for researchers. Separately, Google DeepMind reported that AlphaEvolve sped up a matrix-multiplication component used in Gemini, reducing overall training time by 1%. The examples show possible routes for AI to contribute to its successors, but they do not establish a self-sustaining or accelerating research loop.
METR said on November 5 that Anthropic is likely improving model capabilities on well-specified tasks at about twice its earlier rate, though the organization emphasized high uncertainty and described the work primarily as AI assistance to researchers. The assessment adds evidence to a broader question: whether AI systems can help develop their successors, potentially changing the pace of AI research.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
METR, an independent evaluation organization, said it would not be surprised if the rate were as low as 1.3 times the earlier rate or as high as three times. The assessment concerns improvement on well-specified tasks at Anthropic; it is not a measurement of a fully automated research process. METR redacted much of the evidence it considered most informative, limiting what readers can independently inspect from the published account.
A separate example comes from Google DeepMind’s AlphaEvolve. In a report dated May 14, 2025, the company said the system improved a matrix-multiplication component used in Gemini by 23%, reducing the model’s overall training time by 1%. That is a company-reported engineering result, not evidence that AI independently designed and trained a successor model.
The proposed research loop is straightforward in outline: a model helps researchers propose changes, write code, run experiments and examine results; tests determine which ideas hold up; successful changes are incorporated into a stronger model that can assist with later work. The source essay presents this as a schematic of a possible process, not a measured workflow at any laboratory. Each stage requires reliable evidence and human or institutional decisions about what to pursue and adopt.
How Research Gains Could Compound
If AI assistance makes some research tasks faster or cheaper, laboratories may be able to test more ideas with the same time or computing budget. When improvements reduce resource use at several stages, the savings can add up. The source essay illustrates this with a simplified calculation: three independent 10% savings would leave a job using about 73% of its original compute, allowing roughly 37% more such jobs on the same budget. This is an illustration, not a measured result from a lab.
The broader concern and opportunity is that the research tool could contribute to improving the next version of that tool. If reliable contributions become more extensive, the pace of model development could shift, affecting businesses and people making plans around AI. The cited assessment and optimization are relevant examples, but they do not show how large or durable any future acceleration will be.
Technical capability is only part of the impact. Access to systems, ownership of infrastructure and distribution of benefits will also depend on deployment choices and institutions. A faster research process would not, by itself, determine who can use resulting systems or who gains economically.
Products Worth Considering
Bose SoundLink Max Portable Bluetooth Speaker, IP67 Waterproof and Dustproof, Up to 20 Hours of Battery Life, Black
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black
Bose QuietComfort Wireless Noise-Cancelling Headphones, Cypress Green
As an affiliate, we earn on qualifying purchases.
From AI Assistance to Successors
AI research involves a chain of work: reviewing prior results, inspecting implementations, proposing experiments, writing code, running tests and deciding whether an apparent gain is real. Models may be useful at individual steps, but a convincing result can still be wrong. A code change might accidentally make a test easier, use different computing resources or produce a gain that disappears under a different random seed.
That is why the relevant question is not only whether a model can generate a promising change. It is also whether it can help test the change fairly, find flaws in the evidence and recognize when a result is too weak to support a conclusion. The source describes METR’s assessment as primarily about AI assistance to researchers, a narrower claim than autonomous AI research.
The AlphaEvolve report offers a concrete example of AI-assisted optimization with a measurable effect on part of Gemini’s training process. Its reported 1% reduction in overall training time shows how a localized improvement can matter at scale, while also illustrating the distance between improving one component and automating the full research cycle.
“It “would not be surprised” by 1.3x or 3x.”
— METR
Limits of the Current Evidence
METR described its estimate as highly uncertain and redacted much of the evidence it found most informative. The public account therefore does not let readers fully assess the underlying evidence or determine how broadly the estimate applies beyond the specified tasks at Anthropic.
The cited material does not establish that AI can independently set research priorities, reliably judge results across a whole laboratory, or create successive generations of models with little human direction. The AlphaEvolve result is a company report about a particular optimization; the source does not provide an independent evaluation of that result.
Practical constraints include experiment time and available hardware, transferring results from small tests to large systems, choosing worthwhile research directions, maintaining continuity across work and securing suitable data. The pace of progress may be uneven. How benefits and infrastructure are distributed also remains a question of deployment and institutions, not just model capability.
What Researchers Will Measure
The next useful evidence would show which research tasks AI systems can handle, how much human supervision they need, and whether their contributions survive fair, repeatable tests. Measurements across more tasks and over time could clarify whether assistance is broadening and whether it changes the rate of capability improvement.
For now, the reported developments support a limited conclusion: AI tools are being used to assist research, and at least one company has reported a measurable efficiency gain in a component of Gemini training. Whether these contributions form a dependable loop that materially accelerates model development remains an open question.
Key Questions
What did METR report about Anthropic?
METR said Anthropic is likely improving model capabilities on well-specified tasks at about twice its earlier rate, with high uncertainty. It said the work it described was primarily AI assistance to researchers.
What did AlphaEvolve change?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, cutting overall training time by 1%.
Does this show AI can build its own successor?
No. The cited examples indicate AI assistance in research and a reported optimization. They do not establish a fully autonomous process for designing, testing and training successor models.
Why might small efficiency gains matter?
If improvements reduce the computing or time required for research, a team may be able to run more experiments with the same resources. The source’s example of three 10% savings is a simplified illustration, not a measured outcome at a laboratory.
What could limit AI research acceleration?
Constraints include experiment time, hardware availability, transferring small-scale results to large systems, selecting research directions, continuity and data quality. The pace and distribution of any gains remain uncertain.
Source: Thorsten Meyer AI
