What Could AI Improve About The AI That Follows?
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: What Could AI Improve About The AI That Follows? on ThorstenMeyerAI.com

Age 18–24?Offer from Amazon

Prime made for students and young adults

  • Fast, free delivery for dorm and study essentials
  • Prime Video and Amazon Music included
  • Member-only deals
Try Prime for Young Adults Free trial for eligible 18–24 year olds
As an affiliate, we earn on qualifying purchases.

TL;DR

METR’s November 5 assessment said Anthropic is likely improving capabilities on well-specified tasks at about twice its earlier rate, with high uncertainty, and that AI assistance to researchers was probably the main contributor. A Google DeepMind example shows a narrower efficiency gain in Gemini training. Whether these developments lead to sustained, rapid progress depends on technical and institutional constraints.

AI-assisted research likely helped Anthropic improve model capabilities on well-specified tasks at about twice its earlier rate, according to an assessment published November 5 by independent evaluator METR. The estimate carries high uncertainty, and the report describes AI primarily assisting researchers rather than conducting research autonomously; the finding matters because AI tools may contribute to developing the systems that come after them.

METR assessed information from Anthropic covering August through October 2026. It said it would not be surprised if the rate of capability improvement were as low as 1.3 times the earlier rate or as high as three times. The assessment therefore points to possible acceleration, not a precise or settled measurement. METR judged AI-assisted research likely responsible for most of the acceleration it examined.

The public account has a significant evidentiary limit: Anthropic redacted much of the material METR considered most informative. Readers outside the evaluation cannot independently inspect that underlying evidence from the published description. METR’s conclusion is an assessment based on information it reviewed, not a claim that AI now carries out an entire research program without human direction.

A separate example from Google DeepMind concerns computing efficiency. The company reported that AlphaEvolve 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 specific component and training process; it does not by itself establish a general rate of AI research acceleration.

At a glance
reportWhen: METR assessment published November 5, 2…
The developmentA November 5, 2026, METR assessment found that AI-assisted research likely contributed to faster capability improvements at Anthropic, though the estimate is uncertain and much underlying evidence is redacted.

How Research Gains Could Compound

The potential feedback loop is straightforward: an AI system helps researchers write or inspect code, design experiments, analyze results, or identify errors; researchers validate useful contributions and use them to build a successor; that successor may then assist with further development. The same tool being improved can contribute to the improvement process. If that contribution becomes reliable, research teams could test more ideas or spend less time on routine work.

Small efficiency improvements can also matter when they apply repeatedly. In the source’s simplified illustration, three independent savings of 10% each would leave a task using about 73% of its original compute. With a fixed compute budget, that would allow roughly 37% more of those tasks. This is an illustration, not a forecast: real savings may not be independent, repeatable, or applicable across a full training run.

For businesses and people planning around AI, the distinction matters. If development speeds up, product capabilities and the cost of deploying them may change more quickly than expected. But technical progress alone does not determine who can access the systems, who controls the computing infrastructure, or how economic benefits are distributed. Those outcomes depend on deployment choices and institutions as well as model capabilities.

As an affiliate, we earn on qualifying purchases.

From Research Tasks to Successors

AI research involves more than generating a promising idea. A researcher might review prior work, inspect an implementation, prepare data and configurations, run an experiment, investigate failures, repeat promising tests, and judge whether a result is real. AI can potentially assist at several points, but a plausible-looking result can still be wrong.

For example, a reported improvement might come from an easier test, different computing resources, or a result that disappears with another random seed. Useful assistance must therefore include fair testing and careful interpretation, not just code generation. A system’s contribution is more meaningful when it survives checks and researchers can tell when evidence is too weak to support a conclusion.

The METR assessment describes current work mainly as AI helping researchers. That is different from a system independently choosing research priorities, carrying out experiments, and reliably deciding which findings should shape the next model. The gap between assistance on defined tasks and autonomous research across a full development cycle remains important when interpreting claims about future progress.

“would not be surprised”

— METR, in its November 5, 2026 assessment

Limits of the Acceleration Evidence

The published material does not let outside readers fully evaluate METR’s evidence because much of the material it found most informative was redacted. The assessment gives a wide range rather than a single confident estimate, and the described analysis covers a limited period, August through October 2026. The source material does not provide enough detail here to determine how the rate was calculated or how it compares across other laboratories.

It is also unclear how much assistance will translate into durable gains at larger scales. Work on small experiments may not carry over to training larger systems, while experiment time, access to hardware, data quality, and continuity of research can all constrain progress. Choosing which problem to work on remains another challenge. The available examples do not establish that these constraints have been overcome or that acceleration will continue at the same pace.

Finally, the source’s reference to possible progress toward superintelligence is an interpretation of the feedback-loop risk, not a finding that superintelligence is imminent or inevitable. The assessment concerns research acceleration at Anthropic, and the DeepMind example concerns one reported efficiency improvement. Neither settles what future systems will be capable of or when those capabilities may arrive.

What Future Evaluations Need to Show

The next useful evidence would make it easier to check whether AI contributions produce repeatable improvements: clearer descriptions of the tasks involved, the level of human supervision, independent validation, and how gains hold up across experiments and larger training runs. Further assessments could also show whether reported acceleration continues over longer periods or remains limited to particular tasks.

For now, readers should distinguish AI assistance from autonomous AI research, and specific efficiency results from broad claims about development speed. The next milestone is not established in the source material; no follow-up date or further METR assessment is specified. How quickly research changes, and who benefits from any gains, remain open questions.

Key Questions

What did METR conclude about Anthropic’s AI research?

METR judged that Anthropic was likely improving model capabilities on well-specified tasks at about twice its earlier rate. It said the estimate had high uncertainty and described AI assistance to researchers as the likely main contributor.

Does the assessment show that AI is doing research autonomously?

No. METR’s account describes AI primarily assisting human researchers. It does not establish that AI independently sets research priorities and completes a full development cycle.

How certain is the reported acceleration?

METR described substantial uncertainty. It said it would not be surprised by an estimate ranging from 1.3 times to three times the earlier rate. Much of the evidence it considered most informative was redacted.

What did Google DeepMind report about AlphaEvolve?

Google DeepMind said AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, cutting overall Gemini training time by 1%. This is a company-reported result for a specific component, not a general measurement of AI research acceleration.

Does this mean rapid AI progress is guaranteed?

No. The source describes possible feedback between AI assistance and model development, while identifying constraints including experiment time, hardware, data quality, and the difficulty of choosing research directions. Continued acceleration is not established.

Source: ThorstenMeyerAI.com

You May Also Like

Why Skepticism Faded—and How the Renaissance Revived It

Theories of certainty faded over time, but the Renaissance sparked a revival of skepticism that transformed our approach to knowledge and inquiry.

Biggest Dark Matter Detector Spots A Single Weird Particle

The world’s largest dark matter detector has identified a single unusual particle, raising questions about dark matter and particle physics.

Will It Rain In Oklahoma City On Aug 22, 2026?

Current market activity suggests uncertainty about rain in Oklahoma City on August 22, 2026, with no definitive weather forecast available yet.

How Academic Skeptics Sparred With the Stoics—And Why It Still Matters

Lurking beneath ancient debates between Skeptics and Stoics lies insights that challenge your views on certainty and virtue today.