🔍 Read the full analysis: Could Today’s AI Help Create Tomorrow’s AI? on ThorstenMeyerAI.com
Get smart everyday buys 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’s November 5 assessment said Anthropic is likely improving model capabilities on well-specified tasks at about twice its earlier rate, with substantial uncertainty. It attributed much of the acceleration to AI assistance for researchers, while noting that public evidence is limited; a separate DeepMind example shows a narrower efficiency gain from AI-designed code.
AI assistance may be speeding up AI research at Anthropic, according to an assessment published November 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 stressing substantial uncertainty and describing the work primarily as researchers receiving AI assistance—not autonomous systems building successor models on their own.
METR said its assessment drew on internal information from August through October 2026. Its estimate was not presented as a precise measurement: the organization said it would not be surprised by an acceleration of 1.3 times or 3 times. The range reflects uncertainty about how much the reported changes represent a lasting shift in research productivity.
The public account has a significant evidence gap. Anthropic redacted much of the material METR considered most informative, leaving outsiders with less ability to examine the underlying evidence directly. METR’s conclusion is an evaluator’s judgment based on access to information that has not been fully made public.
A separate example comes from Google DeepMind. The company reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing Gemini’s overall training time by 1%. That is a specific efficiency result, not evidence that the system independently designed a new model or accelerated every part of AI research.
How Research Gains Could Compound
The significance is that AI tools can potentially contribute to the work that improves AI systems. Researchers may use models to inspect code, propose experiments, analyze results or find implementation errors. If those contributions are reliable, a team could investigate more ideas with the same staff and computing resources. A stronger model might then assist with more of that work, creating a feedback loop.
That loop is a possibility, not a confirmed outcome of the METR assessment. Its reported finding concerns faster capability improvement on well-specified tasks and mostly researcher assistance. It does not establish that AI systems can independently set research agendas, validate every result or reliably produce their own successors.
Efficiency gains can still matter even when they appear modest. The source article offers a simplified illustration: if three independent changes each save 10% of a job’s computing needs, the job would use about 73% of its original compute. Under the same fixed budget, that would allow roughly 37% more of those jobs. This is an illustration based on assumed independent savings, not a measured result from Anthropic or DeepMind.
Faster research could affect businesses and workers planning around AI, but technical progress alone does not determine who benefits. Access to systems, ownership of computing infrastructure, deployment choices and institutions will shape how any productivity gains are distributed.
Products Worth Considering
Logitech MX Master 4 Ergonomic Wireless Mouse with Haptic Feedback, Graphite
FlexiSpot E6 Electric Standing Desk, 55 × 28 in, Bamboo Top, White Frame
FLEXISPOT EN1 Electric Standing Desk, 48 x 24 Inches, Black
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Autonomy
AI research involves more than generating an idea. A team might need to read prior work, inspect code, prepare data, configure an experiment, run tests, investigate failures and decide whether a result holds up. AI can assist at several points without taking over the whole process. The distinction matters: helping a researcher complete routine tasks is different from independently choosing and validating a research direction.
For example, a system might report that a code change improved a model’s performance by 5%. That result could be misleading if the test became easier, the runs used different amounts of compute or the gain disappears under another random seed. Reliable research requires checking those possibilities and repeating promising experiments. The source argues that useful AI assistance must include attention to whether results are sound, not just the ability to produce a plausible proposal.
The AlphaEvolve example illustrates a narrower kind of contribution: improving a component of a training process. DeepMind’s reported 23% component-level speedup translated into a 1% reduction in overall training time. The difference shows why local gains should not be confused with equivalent improvements across an entire model-development process.
“It would not be surprised by 1.3x or 3x.”
— METR, as described in its November 5 assessment
Evidence and Limits Still Matter
The METR estimate carries substantial uncertainty, and the public cannot inspect much of the evidence the evaluator found most informative because Anthropic redacted it. The assessment therefore does not provide a fully transparent, independently reproducible account of the underlying work in the source material.
It is also unclear how much of the reported acceleration came from AI assistance compared with other changes in research methods, staffing, compute or project selection. The assessment described AI assistance as likely responsible for most of the current acceleration, but did not establish that AI systems were acting autonomously across the research process.
Several practical constraints could slow progress: experiment time and hardware availability, difficulty transferring results from small trials to large systems, selecting promising research questions, maintaining continuity across long projects and ensuring data quality. The pace may be uneven. The source material does not establish when, or whether, AI research assistance will become sufficiently reliable to produce sustained rapid acceleration.
Watch for Stronger Public Evidence
The next useful signal will be further evidence showing what AI systems actually do in research workflows, how much supervision they require and whether their contributions survive independent checks. Publicly accessible methods and results would help outside researchers assess claims that are now partly obscured by redactions.
Readers should also distinguish improvements in a specific task from broader changes in model-development speed. Follow-up assessments could clarify whether gains persist over time, extend beyond well-specified tasks and translate into measurable reductions in the time or compute needed to develop stronger systems. Until then, the METR finding is evidence of possible acceleration—not a settled forecast of what comes next.
Key Questions
What did METR report about Anthropic?
METR judged that Anthropic was likely improving capabilities on well-specified tasks at about twice its earlier rate. It reported substantial uncertainty and said the work it described was primarily AI assistance to researchers.
Does the assessment show that AI is building AI autonomously?
No. The source describes AI assisting researchers. It does not establish that AI systems independently choose research goals, validate results and build successor models without human oversight.
What did AlphaEvolve change for Gemini?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall training time by 1%. The reported result concerns that specific component and training process.
Why is the METR finding hard to verify publicly?
Anthropic redacted much of the evidence METR considered most informative. The public account therefore offers less direct access to the material behind the evaluator’s judgment.
What could limit AI-driven research acceleration?
Constraints may include experiment time, hardware, the difficulty of scaling small-test results, choosing research directions, project continuity and data quality. The source says the pace of progress may be uneven.
Source: ThorstenMeyerAI.com
