TL;DR
METR’s November 5 assessment said Anthropic is likely improving model capabilities on well-specified tasks at about twice its earlier rate, but stressed that the estimate is highly uncertain and describes mainly AI assistance to researchers. Separately, Google DeepMind reported that AlphaEvolve sped up a matrix-multiplication component used in Gemini by 23%, reducing overall training time by 1%. These examples show ways AI can assist AI development; they do not establish that a self-improving cycle is already operating at a measured pace.
METR said on November 5 that Anthropic is likely improving its models’ capabilities on well-specified tasks at around twice its earlier rate, with substantial uncertainty around that estimate. The independent evaluation organization described the work as primarily AI assistance to researchers, a development that matters because AI tools could contribute to the research that produces future models.
Get everyday essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
METR assessed evidence about AI research acceleration at Anthropic. Its estimate is not a precise measurement: the organization said it “would not be surprised” by a rate anywhere from 1.3 times to three times the earlier pace. METR also redacted much of the evidence it considered most informative, limiting what readers can independently examine from the material summarized here.
A separate example comes from Google DeepMind’s AlphaEvolve. DeepMind reported that the system made a matrix-multiplication component used in Gemini 23% faster. The company said that change reduced Gemini’s overall training time by 1%. The component-level improvement and the overall training-time reduction refer to different scopes; the larger component gain did not translate into an equivalent reduction across the full training run.
Together, the examples point to ways AI can assist parts of research and computing work. They do not show that a model independently designed and built its own successor, or establish a measured, repeating cycle in which each new model accelerates the next one. The loop described in the source material is a schematic explanation of how such a process might work, not a documented process at either lab.
How Research Gains Could Compound
If AI systems help researchers run experiments, improve code, or find errors, even modest gains in individual tasks could affect how much work a research team completes. The potential significance lies in the possibility that a more capable model could help with more of the work behind its successor. That could make the pace of development partly dependent on tools built by earlier rounds of development.
The source offers a simplified illustration: three independent savings of 10% would leave a job requiring about 73% of its original computing resources. With a fixed compute budget, that would allow roughly 37% more of those jobs. This is an illustration based on independent savings, not a reported result from a lab. It does not account for whether improvements would be independent or apply to the same workload.
For companies, workers, and public institutions, the pace matters because plans based on gradual improvement may be strained if research capabilities rise more quickly. But technical gains alone do not determine who can use the systems, who controls the infrastructure, or how the benefits are shared. Those outcomes also depend on deployment choices and institutions.
Products Worth Considering
DJI Osmo 360 Motorcycle Handlebar POV Combo
Samsung Odyssey G55C 32-Inch QHD 1000R Curved Gaming Monitor
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black
As an affiliate, we earn on qualifying purchases.
From Research Help to Successors
The proposed process has several stages: a model helps with research, researchers test its suggestions, useful results are retained, and a stronger model joins the work. Assistance might include proposing changes, writing code, running experiments, or checking results. Each stage has ways to fail. A change can appear successful because a test was easier, computing budgets differed, or results depended on a particular random seed.
That is why fair testing and human judgment remain relevant in the account provided. A useful contribution has to survive checks that show the result holds up; a plausible suggestion or an impressive initial run is not enough. METR’s assessment concerns acceleration in improving capabilities on well-specified tasks, and its account says the work was primarily AI assistance to researchers. It should not be read as proof that every part of AI research has been automated.
The AlphaEvolve example shows that an AI system can contribute to an optimization with a reported effect on a component used in a major model’s training. The result dates to May 14, 2025. It predates METR’s November 2026 assessment and is a separate example, not evidence that AlphaEvolve caused the rate METR estimated at Anthropic.
“AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, cutting overall training time by 1%.”
— Google DeepMind
Limits of the Acceleration Evidence
METR’s estimate carries high uncertainty, and its redactions mean readers cannot see much of the evidence the organization found most informative. The available account does not specify the underlying task set, the full comparison method, or how the estimate might vary across research activities. It also does not establish that the estimated rate will continue.
More broadly, the source identifies constraints that could slow a research feedback loop: experiment time and hardware availability, differences between small experiments and large systems, decisions about which research directions to pursue, continuity of work, and data quality. The pace of progress could be uneven. The material does not quantify how much each constraint affects current development or when any might become limiting.
It remains unclear how much of AI research can be performed reliably with little supervision, how often proposed improvements survive independent testing, and whether gains across separate tasks can be combined. The examples here do not settle those questions or establish a timeline for superintelligence.
What Future Assessments Must Show
The next useful evidence would show how AI assistance performs across a defined set of research tasks, how results are checked, and how the pace compares with an explicit earlier baseline over a stated period. Further detail about METR’s methods and the evidence withheld from its assessment could help readers judge the estimate’s scope. Follow-up evaluations will also show whether the reported acceleration persists or changes.
For claims about a self-improving cycle, the key development to watch is evidence that models contribute reliably to research on stronger successors, with improvements verified under fair tests. Until those results are available, the reported findings support a narrower conclusion: AI is assisting some research work, and at least one reported optimization improved a component used in Gemini’s training. How quickly that assistance expands remains unresolved.
Key Questions
What did METR report about Anthropic?
METR estimated that Anthropic is likely improving model capabilities on well-specified tasks at around twice its earlier rate. It described the estimate as highly uncertain, with a plausible range it would not be surprised to see from 1.3 times to three times.
Did METR say Anthropic’s models are building their own successors?
No such finding is described in the source material. METR characterized the work it assessed as primarily AI assistance to researchers. The repeating successor process is presented as a schematic of a possible feedback loop, not a measured process at Anthropic.
What did AlphaEvolve improve?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing Gemini’s overall training time by 1%.
What remains uncertain about AI-assisted research?
The available account does not establish how broadly AI assistance works across research tasks, whether improvements will persist, or how quickly models might contribute to building stronger successors. METR also redacted much of the evidence it considered most informative.
Source: Thorsten Meyer AI
