What It Means When AI Starts Improving AI
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: What It Means When AI Starts Improving AI on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the little things that make your day delivered free with Prime

  • Fast, free delivery on millions of items
  • Prime Video, Amazon Music and more included
  • Member-only deals all year
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

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. The report attributed much of the acceleration to AI assistance for researchers, while noting that much of the supporting evidence was redacted. The finding points to a possible feedback loop, but does not establish that AI research is fully automated or that progress will accelerate without limit.

METR said on November 5 that Anthropic was likely improving its models’ capabilities on well-specified tasks at about twice its earlier rate, with high uncertainty. The independent evaluation organization judged that AI assistance to researchers was likely responsible for much of the acceleration it assessed, a finding that adds evidence to the possibility of a feedback loop in which AI systems help develop more capable successors.

METR’s assessment drew on internal information from August through October 2026. The organization described its estimate cautiously: it said it would not be surprised by an acceleration as low as 1.3 times or as high as three times. The report concerned improvements on well-specified tasks, not a general measurement showing that every aspect of AI development had doubled in speed.

The source article says the work METR described was primarily AI assistance to human researchers. That distinction matters: the assessment does not establish that AI systems independently ran the full research process, selected all research priorities, or built a successor model without human involvement. It indicates that assistance may already be contributing to the pace of model-capability improvements.

A separate example cited in the article comes from Google DeepMind, which reported that its AlphaEvolve system made a matrix-multiplication component used in Gemini 23% faster. DeepMind said the change reduced Gemini’s overall training time by 1%. This is a company-reported engineering result, not evidence that all AI research is accelerating at the same rate. It illustrates how a narrow improvement may have wider effects when it is used in a larger training process.

At a glance
reportWhen: METR published its assessment November…
The developmentMETR published an assessment linking AI assistance to a likely acceleration in Anthropic’s model research, while emphasizing substantial uncertainty.

How Research Gains Could Compound

The potential significance is a change in the pace of producing AI systems, not just an improvement in one model. If AI tools help researchers write code, examine results, identify errors, or test approaches, a team might investigate more ideas with the same time and resources. If a resulting model then provides more effective research assistance, some of those gains could feed into later development.

The source article gives a simplified illustration: three independent savings of 10% would leave a job requiring about 73% of its original compute. Under that illustration’s assumptions, a fixed compute budget could run roughly 37% more of those jobs. This is arithmetic, not a measured forecast of real-world research output; actual gains depend on whether savings are independent and whether the experiments remain comparable and useful.

For businesses, workers, and governments planning around AI, the question is whether progress could become less predictable if research itself becomes more productive. The article argues that planning should allow for changing rates of improvement. But the METR estimate is uncertain and limited in scope, so it does not establish a specific timeline for future capabilities or their effects on employment and the wider economy.

Products Worth Considering

Therabody Theragun Mini 3rd Generation Portable Massage Gun, Black

Therabody Theragun Mini 3rd Generation Portable Massage Gun, Black

Hyperice Hypervolt Go 3 Percussion Massage Gun with Carry Case

Hyperice Hypervolt Go 3 Percussion Massage Gun with Carry Case

As an affiliate, we earn on qualifying purchases.

From Research Assistance to Feedback

AI research involves more than generating a promising idea. A team must read relevant work, inspect code, prepare data and experiments, measure results, investigate failures, repeat promising tests, and decide whether an apparent gain is real. AI assistance can apply at several points in that process, but reliable evaluation remains necessary. A system might report a 5% improvement because it inadvertently changed the test conditions, used a different amount of compute, or benefited from a particular random seed.

The source article’s central idea is that the tool being improved may also contribute to the work of improving it. That does not mean the tool can already perform every research task. The relevant questions include how much of the sequence it can handle, how much supervision it needs, and whether its contributions hold up under independent checks.

METR’s assessment is one recent piece of evidence, while the AlphaEvolve example concerns a specific optimization reported by Google DeepMind. Together they illustrate different ways AI may assist development: supporting researchers and improving a component of a training system. Neither example alone establishes a general, self-sustaining cycle of rapid progress.

“It would not be surprised by 1.3x or 3x.”

— METR, as described in its November 5 assessment

Limits of the Current Evidence

The assessment’s estimate is explicitly uncertain, and much of the evidence METR considered most informative was redacted by Anthropic, according to the source article. That limits what outside readers can independently examine about the methods and underlying observations. The article does not provide enough detail to resolve how the estimate was calculated or how representative the assessed tasks were of AI research as a whole.

It is also unclear how much of the reported acceleration came from AI assistance compared with other changes in research methods, staffing, infrastructure, or compute. The source says METR judged AI assistance likely responsible for most of the current acceleration it described, but that remains an assessment with substantial uncertainty, not a definitive causal finding.

More broadly, the process could slow down. The article identifies constraints including experiment time and hardware, difficulty carrying results from small tests to large systems, decisions about which research directions to pursue, continuity, and data quality. The evidence presented does not show that these constraints have disappeared or that AI assistance will keep increasing productivity at a steady rate.

What Researchers Need to Track

The next useful evidence would include clearer, independently checkable measurements of how AI assistance changes research output over time: what tasks are being automated, how much supervision is required, and whether results survive repeated tests and independent review. Further assessments could also clarify whether gains apply beyond well-specified tasks and whether AI assistance is responsible for them rather than merely accompanying other changes.

For now, the source article recommends allowing for the possibility that the pace of improvement may change, while recognizing that development is likely to be uneven. How the effects are distributed will depend not only on technical capability but also on who can use the systems, who controls the infrastructure, and how benefits are shared. The available assessment does not settle those questions or provide a firm timetable for what comes next.

Key Questions

What did METR report about Anthropic?

METR judged that Anthropic was likely improving model capabilities on well-specified tasks at about twice its earlier rate. It gave a wide uncertainty range and said AI assistance to researchers was likely responsible for much of the acceleration it assessed.

Does the assessment show that AI is independently building better AI?

No. The source describes the work assessed as primarily AI assistance to human researchers. It does not establish that AI systems independently managed the full process of developing successor models.

How strong is the evidence?

METR’s estimate carries substantial uncertainty, and the source article says Anthropic redacted much of the evidence METR found most informative. That makes the finding relevant but difficult for outsiders to verify in full.

What is the AlphaEvolve example?

Google DeepMind reported that AlphaEvolve made a matrix-multiplication component used in Gemini 23% faster, reducing overall Gemini training time by 1%. It is a specific company-reported optimization, not a general measure of AI research acceleration.

What could slow AI-assisted research?

The source identifies experiment time and hardware limits, difficulties scaling small experiments, research-direction choices, continuity, and data quality as possible constraints. It does not say how strongly each currently limits progress.

Source: ThorstenMeyerAI.com

You May Also Like

The Skills Marketplace, Six Months Later: Predicted vs Actual

An analysis of the skills marketplace’s growth, structure, and challenges six months after predictions, highlighting confirmed developments and ongoing uncertainties.

The United States: The High-Variance Bet

The United States is pursuing a minimal regulation strategy for AI, emphasizing market dynamism over government oversight, with significant local experiments on social support.

Why Europe’s Sovereign AI Is Largely Canadian In Nature

Analysis of the recent acquisition of German AI firm Aleph Alpha by Canadian company Cohere reveals complex ownership, infrastructure, and strategic implications for European AI sovereignty.

Prime Big Deal Days Guide To AI Automation Tools For Small Businesses

AI automation tools for small businesses now run $15–50/month. Here’s what to buy during Prime Big Deal Days and what to skip.