🔍 Read the full analysis: The Possibilities When AI Helps Improve AI on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
METR’s November 5 assessment said Anthropic was likely improving model capabilities on well-specified tasks at about twice its earlier rate, with high uncertainty and AI assistance likely responsible for most of the acceleration. A separate Google DeepMind example showed a system optimizing part of Gemini’s training process, though the reported overall training-time reduction was 1%. These examples suggest a possible feedback loop, but its speed and eventual scale remain unclear.
AI assistance may be accelerating some AI research, according to an independent METR assessment of work at Anthropic published November 5. METR judged that Anthropic was likely improving model capabilities on well-specified tasks at roughly twice its earlier rate, while stressing substantial uncertainty; the assessment said the work it described was primarily AI assistance to researchers, not fully autonomous research.
METR said its assessment drew on internal information from August through October 2026. Its estimate was not a precise measurement: the organization said it would not be surprised by an acceleration as low as 1.3 times or as high as three times. Anthropic redacted much of the evidence METR considered most informative, limiting what outside readers can independently inspect.
A separate example comes from Google DeepMind, which reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster. DeepMind said the change reduced Gemini’s overall training time by 1%. The example illustrates how a targeted improvement to one part of a large system can have a smaller effect on the full process.
The broader possibility is a feedback loop: AI tools help researchers write or inspect code, propose experiments, analyze results, or identify errors; useful, independently checked contributions then help produce better models. Those models could assist with later research. The source article presents this as a reason to take rapid progress seriously, not as proof that AI can independently direct or complete the research process.
How Research Gains Could Compound
If AI assistance lets a research team test more ideas or spend less time on routine coding and debugging, the team could make more progress with the same staff and computing budget. If a stronger model then helps with subsequent research, improvements could feed into later development. That possibility matters to companies, workers, and public institutions planning around AI because the pace of progress could change, rather than follow a steady path.
The gains need not be dramatic in any single step to matter. The source article offers a simplified illustration: three independent 10% savings would leave a task needing about 73% of its original computing, allowing a fixed budget to run roughly 37% more of those jobs. This is an illustration, not a measured result from either METR or DeepMind; actual gains depend on whether the savings are independent and apply to comparable work.
Technical capability alone will not determine who benefits. Access to AI tools, ownership of computing infrastructure, and decisions about how gains are shared will also shape the effects. The evidence cited so far concerns research assistance and a specific training optimization, not a settled forecast of broad labor-market change.
Popular right now

Ninja Air Fryer XL, 6.5 Qt, 6-in-1, MaxCrisp, Dark Grey (AF181)

Sony WH-1000XM6 Wireless Noise-Canceling Headphones, Black

iRobot Roomba 105 Robot Vacuum with LiDAR Navigation
As an affiliate, we earn on qualifying purchases.
From Research Assistance to Successors
AI research involves more than generating a promising idea. A team must review relevant work, inspect existing code, design an experiment, prepare data and measurements, investigate failures, repeat promising tests, and decide whether an apparent improvement is real. AI can assist at several points, but a result that looks good may be misleading if tests are inconsistent, computing budgets differ, or an improvement disappears under another random seed.
That makes reliable evaluation and human oversight central to the question of research acceleration. A useful system must do more than produce code or report a gain: its contributions need to withstand fair testing and independent checks. The METR assessment described mostly AI assistance to researchers, while the DeepMind example concerned a specific component of Gemini’s training process. Neither example establishes that a system can autonomously choose research goals and validate a successor model end to end.
What the Evidence Cannot Establish
The METR estimate carries substantial uncertainty, and outsiders cannot review much of the evidence METR said was most informative because Anthropic redacted it. The assessment concerns a particular organization and period; the source material does not establish that the same pace applies across AI labs or will continue.
It is also unclear how much of the reported acceleration comes from AI assistance rather than other changes, or how much current assistance can scale into systems that select research directions and reliably judge results. Experiment time, hardware availability, the gap between small tests and large-scale systems, continuity of work, and data quality could all slow progress. The timing and extent of any broader economic effects remain unknown.
Testing Whether the Loop Persists
The next useful evidence will be repeated, transparent evaluations showing how AI assistance changes research output over time, what tasks models can perform, how much human supervision they require, and whether gains hold up in independent tests. Further reporting on the METR assessment’s methods and any additional evidence made available would help outsiders judge its estimate.
Researchers and decision-makers will also need to track practical limits, including computing capacity and whether improvements found in small experiments carry over to large systems. For now, the documented examples support a measured conclusion: AI is helping with parts of AI research and development, but the pace, reliability, and eventual reach of that assistance remain unsettled.
Key Questions
What did METR conclude about Anthropic’s AI research progress?
METR judged that Anthropic was likely improving model capabilities on well-specified tasks at roughly twice its earlier rate. It emphasized high uncertainty, including a range it said could plausibly be 1.3 to three times.
Does the assessment show that AI is conducting research autonomously?
No. The source material says the work METR described was primarily AI assistance to researchers. It does not establish end-to-end autonomous research.
What did AlphaEvolve improve?
Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, cutting overall training time by 1%.
Why is the METR finding difficult to verify independently?
Anthropic redacted much of the evidence METR considered most informative. That leaves external readers with limited access to the underlying information behind the assessment.
What could slow AI-assisted research?
Potential constraints include experiment time, hardware, the difficulty of transferring results from small tests to large systems, choosing useful research directions, continuity, and data quality.
Source: ThorstenMeyerAI.com
