🔍 Read the full analysis: When The Tool Helps Create Its Successor on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
METR’s November 5 assessment judged that Anthropic’s model capabilities on well-specified tasks are improving at about twice their earlier rate, with high uncertainty. It attributed much of the acceleration to AI assistance for researchers, not fully autonomous AI research. A separate Google DeepMind example shows a targeted optimization can reduce training time, but the broader pace and effects of these gains remain uncertain.
Independent evaluator METR said on November 5 that Anthropic is likely improving model capabilities on well-specified tasks at about twice its earlier rate, with substantial uncertainty around that estimate. METR said the work it assessed was primarily AI assistance to researchers—evidence that AI tools may help speed the development of successor systems, not proof that models are independently conducting AI research.
METR’s assessment was based on internal information from August through October 2026. The organization described its estimate as uncertain: it said it would not be surprised by a rate of roughly 1.3 times or three times the earlier pace. That range signals meaningful uncertainty, rather than a settled measurement of how much AI contributes to research progress.
The public account has an important limitation. According to the source material, Anthropic redacted much of the evidence METR considered most informative. Readers therefore cannot independently inspect all the underlying material, and the public must rely substantially on the evaluators’ interpretation. METR’s account described AI as assisting researchers; it does not establish that an AI system chose and completed the full research agenda without human involvement.
A separate example reported by Google DeepMind illustrates how a narrow efficiency improvement can matter at scale. The company said its AlphaEvolve system made a matrix-multiplication component used in Gemini 23% faster, reducing Gemini’s overall training time by 1%. This is a company-reported result about a particular component and training process, not a general measure of AI research acceleration.
Why Faster Research Could Compound
The potential significance lies in a feedback loop: AI systems can assist work that improves the models and infrastructure used to build later systems. If that assistance reliably helps researchers test more ideas, spot errors, or reduce the resources required for experiments, development could speed up even without full automation. A more capable successor could then provide more useful assistance, although the size and reliability of that effect are not established by the cited examples.
Efficiency gains can also add up. The source offers a simplified illustration: if three independent changes each save 10% of a job’s compute, the work would use about 73% of the original resources. Under that illustrative assumption, the same compute budget could run roughly 37% more jobs. This is an arithmetic example, not evidence that three such savings have occurred together or that real-world improvements will be independent.
For organizations planning AI investment, the distinction between assistance and autonomy matters. Today’s reported research help could raise productivity within existing teams, while broader autonomous research would require systems to select worthwhile questions, run sound experiments, interpret results, and recognize weak evidence. How quickly those capabilities develop could affect the cost and timing of AI deployment, but the available material does not set a reliable timetable.
Popular right now
Dyson Ball Animal 2 Upright Corded Vacuum Cleaner, Fuchsia
Hisense 43-Inch E6 Cinema Series Hi-QLED 4K UHD Smart Fire TV (43E6QF)
Ninja AF101 4-Quart Air Fryer, 4-in-1, Grey
As an affiliate, we earn on qualifying purchases.
From Research Tasks to Successors
AI research involves many linked tasks: reviewing prior work, inspecting code, preparing data, setting up experiments, comparing results, investigating failures, and deciding whether an apparent gain is real. AI can potentially help with individual steps while researchers retain responsibility for direction and validation. A system that produces a promising result is not necessarily producing a sound one; differences in compute, test design, or random variation can make a result look better than it is.
The source’s central argument is that the tool being improved may also contribute to the work of improving it. The METR assessment offers a recent, specific indication of AI assistance within research at Anthropic, while the AlphaEvolve example concerns an optimization in a component used for Gemini training. These are distinct examples and should not be treated as a single demonstration of an autonomous self-improvement cycle.
The broader case also depends on constraints. Experiments take time and hardware, and gains on small tests may not carry over to large systems. Researchers still need to choose promising directions, maintain continuity across work, and use data that supports meaningful conclusions. These limits can make progress uneven even if AI assistance becomes more capable.
“It would not be surprised by 1.3x or 3x.”
— METR, as characterized in its November 5 assessment
Limits of the Public Evidence
The public material does not provide all the evidence METR reviewed because, according to the source, Anthropic redacted much of the most informative material. The precise methods, comparison details, and contribution of different forms of AI assistance therefore cannot be fully assessed from the account provided. METR’s estimate is explicitly uncertain, and its range should not be read as a precise forecast.
It is also unclear how much of the reported acceleration would persist across other laboratories, research tasks, or larger systems. The source gives no basis for concluding that AI is already autonomously carrying out the full cycle of AI research or that faster progress will continue at a steady rate. The AlphaEvolve figures are attributed to Google DeepMind, and the material does not provide an independent evaluation of that result.
Finally, faster technical progress would not by itself determine who can use advanced systems, who owns the computing infrastructure, or how economic benefits are distributed. Those outcomes depend on deployment choices and institutions as well as technical capability.
Testing the Research Acceleration
The immediate next step for readers and evaluators is to look for more public evidence that clarifies how AI assistance is measured: what tasks are included, what comparison period is used, how researchers validate results, and how much human supervision remains necessary. Further assessments could show whether the reported pattern is sustained, limited to particular tasks, or dependent on specific laboratory conditions.
For research teams, the key test is whether proposed improvements survive independent checks and transfer from small experiments to large training runs. That means comparing like with like, repeating promising results, and checking whether efficiency gains remain useful when applied to real systems. The source identifies experiment time, hardware, research direction, continuity, and data quality as possible brakes, but does not say how quickly any of them will constrain progress.
No specific next report or milestone is identified in the supplied material. The pace of change, degree of automation, and eventual effects on work remain open questions; the current evidence supports monitoring the trend, not treating rapid autonomous self-improvement as an established outcome.
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 described the estimate as highly uncertain and said the work it assessed was primarily AI assistance to researchers.
Does the assessment show that AI is doing research autonomously?
No. The source says METR described the work as primarily assistance to researchers. It does not establish that AI systems independently choose research directions and complete the full process without human oversight.
What did AlphaEvolve reportedly improve?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall Gemini training time by 1%. These are company-reported figures for a specific application.
Why is the METR estimate uncertain?
METR described a broad range of possible acceleration, and the source says Anthropic redacted much of the evidence METR considered most informative. The public material does not allow readers to examine all the underlying details.
What could slow AI-assisted research?
The source identifies experiment time and hardware, difficulty transferring small-test results to large systems, research direction, continuity, and data quality as potential constraints. It does not establish which will be the strongest limit or when they might bind.
Source: ThorstenMeyerAI.com
