AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: When AI Starts Improving AI, What Changes? on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

METR’s November 5 assessment estimated that Anthropic’s capabilities on well-specified tasks were improving at about twice the earlier rate, with substantial uncertainty and AI assistance to researchers as the main explanation. Separately, Google DeepMind said AlphaEvolve sped up a matrix-multiplication component used in Gemini by 23%, reducing overall training time by 1%. The evidence points to AI contributing to AI development, but does not establish fully autonomous research or a predictable rate of future progress.

AI systems are beginning to contribute to the work of improving AI, according to a November assessment by the independent evaluation organization METR and a separate engineering report from Google DeepMind. METR estimated that Anthropic’s capabilities on well-specified tasks were improving at roughly twice the earlier rate, while stressing substantial uncertainty; Google DeepMind said its AlphaEvolve system made a component used in Gemini’s training 23% faster, cutting overall training time by 1%.

METR published its assessment on November 5, 2026, drawing on internal information from Anthropic covering August through October. It judged that the company was likely improving model capabilities on well-specified tasks at around twice the previous rate. METR said it would not be surprised by estimates ranging from 1.3 times to three times the earlier rate, indicating that the central estimate is not precise. It attributed most of the acceleration it described to AI assistance to researchers, rather than to fully automated AI research.

The public record has limits: Anthropic redacted much of the evidence METR considered most informative. Readers therefore cannot independently inspect all the information behind the assessment. METR’s conclusion is an evaluator’s judgment based on that work, not a direct measurement that proves AI systems can independently conduct an entire research program.

In a separate example, Google DeepMind reported that AlphaEvolve improved the speed of a matrix-multiplication component used in Gemini by 23%. The company said that change reduced Gemini’s overall training time by 1%. Those figures describe a component-level gain and its reported effect on the broader training run; they do not show that the model itself became 23% more capable.

At a glance
reportWhen: METR assessment published November 5, 2…
The developmentA November 2026 METR assessment and a Google DeepMind engineering report offer concrete, but limited, examples of AI systems helping accelerate AI research and training.

How Research Gains Could Compound

The importance of AI assistance in AI development is not limited to whether a system can replace a researcher. If a model helps people test more ideas, inspect code, diagnose failed experiments or catch flaws in a result, a research team may get more useful work from the same staff and computing budget. The METR assessment suggests that this kind of assistance may already be affecting the pace of development, though the size of the effect remains uncertain.

Efficiency gains can also accumulate. In a simplified illustration, three independent reductions of 10% in computing required for a job would leave it using about 73% of the original compute. With a fixed budget, that could allow roughly 37% more of those jobs to run. This is an illustration, not a measured outcome from either report; actual savings depend on whether improvements are independent and apply to the same work.

Faster research could affect how quickly companies release models and how businesses, workers and governments plan around them. But technical progress alone does not determine who can use the systems, who owns the computing infrastructure or how economic benefits are distributed. Those outcomes also depend on deployment decisions and institutions.

Products Worth Considering

Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black

Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black

Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID (Midnight)

Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID (Midnight)

Amazon Kindle Paperwhite Signature Edition (2024), 32 GB, Metallic Black

Amazon Kindle Paperwhite Signature Edition (2024), 32 GB, Metallic Black

As an affiliate, we earn on qualifying purchases.

From Research Assistance to Automation

AI research involves more than generating a promising idea. A team must review relevant work, inspect existing code, design an experiment, prepare data, run tests and determine whether the results hold up. An apparent improvement may come from a flawed comparison, a changed test, unequal computing resources or a result that disappears with a different random seed.

That makes verification as important as proposing changes. A useful system must help produce results that survive fair testing, not merely report a promising number. Researchers can already use AI for parts of this process, but the examples cited here do not establish that systems can reliably choose research goals, carry out every step without supervision and validate their own conclusions.

The two reports measure different things. METR assessed the pace of capability improvement at Anthropic and offered a judgment about AI’s contribution to research. Google DeepMind described a specific optimization affecting training efficiency. Together, they provide examples of AI contributing to development, but they are not a single standardized measure of how much faster the field is progressing.

“The work it described was primarily AI assistance to researchers.”

— METR, describing the work behind the assessment

Limits in the Public Evidence

The size and durability of the acceleration remain uncertain. METR’s estimate relies on internal information from Anthropic, and the company redacted much of what the evaluator considered most informative. The assessment gives a broad range rather than a precise, independently reproducible public measurement.

It is also not clear how much of the reported pace change came from AI assistance compared with other factors, or whether the gains will continue. AI help can produce misleading results if experiments are poorly controlled; reliable research still requires checks on code, data, compute use and repeatability.

More broadly, the reports do not show that an AI system can independently set research priorities, manage the full research process or develop a successor without meaningful human oversight. Limits include experiment time and available hardware, difficulty applying small-scale results to large systems, research direction, continuity and data quality. How quickly these constraints may ease is not established by the cited evidence.

What Researchers Will Test Next

The next useful evidence will show whether AI-assisted improvements can be independently checked, reproduced and applied across more than one task or model. Further reporting on evaluation methods and the underlying measurements could help clarify how METR arrived at its estimate and how much the pace changed over time.

For engineering claims such as AlphaEvolve’s, the key follow-up is whether the reported efficiency gains persist in routine training and whether comparable improvements can be repeated in other parts of the process. The 1% overall training-time reduction is a specific result; it does not by itself establish a general trend.

For now, the evidence supports a measured conclusion: AI is helping with some work involved in building AI, and those contributions may affect development speed. How far that assistance can extend—and whether it leads to sustained, compounding acceleration—remains an open question.

Key Questions

What did METR estimate about Anthropic’s AI progress?

METR judged that Anthropic’s capabilities on well-specified tasks were improving at around twice the earlier rate. It emphasized substantial uncertainty, saying estimates from 1.3 to three times would not surprise it.

Does the assessment show AI is conducting research independently?

No. METR described the work as primarily AI assistance to researchers. The assessment does not establish that AI systems can independently plan, run and validate an entire research program.

What did AlphaEvolve change for Gemini?

Google DeepMind said AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall training time by 1%. The figures concern efficiency, not a 23% increase in model capability.

Why is the METR estimate uncertain?

METR gave a wide range around its estimate, and Anthropic redacted much of the evidence METR found most informative. That limits what outsiders can independently examine about the assessment.

What could slow AI-assisted research?

Potential constraints include experiment time and hardware, difficulty carrying small-scale results over to large systems, choosing research priorities, maintaining continuity and obtaining high-quality data. The evidence cited does not establish how quickly these limits will change.

Source: ThorstenMeyerAI.com

You May Also Like

Never Trust Facebook: 3 Ways To Protect Your Privacy When Using Meta AI

As interest in Meta AI grows, experts warn about privacy risks; here are three confirmed ways to safeguard your data when using Meta’s AI tools.

EFF To Courts: Don’t Rewrite Copyright Over AI Hype

Electronic Frontier Foundation warns courts against expanding copyright laws due to AI hype, emphasizing the need for clear legal boundaries.

How to Think About Fair Use in Research-Heavy Content

By understanding fair use principles in research, you can navigate complex legal boundaries and ensure your work remains compliant and ethically sound.

The Ethics of Content Scraping and Reposting

Discover the ethical dilemmas of content scraping and reposting, and learn why understanding these boundaries is crucial for responsible online behavior.