AI Accelerates Its Own R&D at Anthropic, Meter Report Highlights

•

 20 min video

•

 6 min read

YouTube video ID: lVbQqjAJpkE

Source: YouTube video by Matthew Berman — Watch original video

PDF

Recursive self-improvement in AI has already begun, according to a report from Meter, a third-party evaluator that had significant access to Anthropic's internal systems, researchers, and data. This assessment confirms what many in the AI community have feared and anticipated, indicating a rapid acceleration in AI research and development primarily driven by AI itself.

The Genesis of the Evaluation

The initiative for third-party evaluation arose a few months prior when leaders of major AI labs, including Anthropic and OpenAI, expressed concerns about the pace of frontier AI development, particularly regarding recursive self-improvement and the eventual emergence of superintelligence. They pledged to grant third-party evaluators "employee-like access" to their internal systems, models, employees, and chat histories.

Meter, a respected independent evaluation company, subsequently collaborated with Anthropic, gaining extensive internal access. The initial findings from this collaboration are now public, revealing significant acceleration in AI capabilities.

Key Findings: AI Accelerating AI R&D

The report, titled "Independent Assessment of AI R&D Acceleration at Anthropic in October 2026," sought to identify signs of AI improving itself and accelerating AI research and development. The evaluators found conclusive evidence of this phenomenon.

It's important to note that while Anthropic had initially promised "employee-like access," this specific Meter evaluation was not part of that broader program. Anthropic retained the right to approve or deny access to information and to determine what findings were published. Despite these limitations, the findings are considered remarkable.

Meter staff now believe that as of October 2026, Anthropic is likely improving the capabilities of its models on well-specified tasks at approximately twice the rate observed between March 2024 and October 2025. This acceleration is not attributed to increased hiring, better human researchers, or more compute power. Instead, Meter was able to isolate AI progress itself as the primary cause.

The report highlights a significant inflection point around late 2025, coinciding with the release of Opus, when models became exceptionally proficient at writing code. While initial acceleration at that time might have been due to non-AI factors (more compute, hiring, human breakthroughs), Meter estimates that by October 2026, approximately 80% of the current acceleration is attributable to AI.

Visualizing the Acceleration

A conceptual chart illustrates this acceleration: * Blue Line (Dotted): Represents historical, linear improvement in model capability. * Gray Line (AECI without AI): Anthropic's internal measure of model capabilities without AI's influence. * Orange Line (AECI with AI): Shows the excess growth due to AI, with a clear inflection point at the end of last year.

This inflection point aligns with the experience of many developers who saw their work fundamentally change due to the advanced code-writing capabilities of models like Opus 4.5.

The Compounding Effect and Intelligence Explosion

A major concern highlighted in the report is the compounding nature of this acceleration. As models improve, they become better at discovering ways to improve themselves more quickly. They also learn to utilize limited compute capacity more efficiently, amplifying human researcher output. This self-reinforcing cycle is reminiscent of the "intelligence explosion" concept.

Anthropic's S1 filing (for its imminent IPO) revealed a tremendous increase in compute utilization efficiency. For example, compute efficiency was 67x in Q4 2025, reached breakeven in Q1 2026, and then doubled to 2x in Q2 2026. This efficiency gain allows for more research and development, creating a positive feedback loop.

This phenomenon echoes Google's Alpha Evolve paper, where AI improved Google's hardware architecture, leading to a 1% efficiency gain—a massive saving at Google's scale.

No Clear Bottlenecks Identified

Perhaps the most striking finding from Meter is the absence of any clear bottlenecks that would impede further growth or lead to diminishing returns. The report states: "We did not identify any clear bottlenecks, limited inputs that would prevent further growth or steep diminishing returns to further scaling that would get in the way of such compounding acceleration." This suggests an unimpeded path toward further AI acceleration.

The report further notes that Anthropic appears to have made as much progress on benchmarks in the past year as it did between October 2023 and October 2025, underscoring the accelerating pace. Internal model capability evaluations (AECI) show a largely linear growth until late 2025, followed by a sharp, accelerating curve after the release of Opus 4.5 and subsequent models like Mythos. The steepness of this curve is itself accelerating.

Meter received convincing evidence that this acceleration was not due to an increase in training compute scale, although Anthropic redacted further details on this evidence. This is significant, as Anthropic has historically faced compute limitations. The efficiency gains are attributed to AI R&D.

Employee Perspectives and AI's Contribution

Meter's evaluation included interviews with Anthropic employees, who reported significant productivity uplifts due to AI. Evidence was gathered from: * Interviews * Self-reported uplift * Headcount * Lines of code * Experiment usage * Qualitative reflections from leadership

Quotes from researchers included: * "Could definitely endorse a greater than 2x speed up." * "My inside view is by far the largest effect is model acceleration. The AI acceleration is larger than all the other effects." * "Now only bottlenecked on ideas and compute." (Though Meter's findings suggest compute is not a bottleneck, the "ideas" bottleneck is interesting, implying human intuition might still be crucial for promising research directions, at least for now.) * "I would definitely endorse greater than 2 years of progress in one year and I don't think four is crazy."

While AI is not yet making fully autonomous contributions to the most critical parts of the model training process, researchers described "significant signs of life" suggesting such contributions could occur soon. This aligns with recent OpenAI research on mathematical proofs, indicating AI's ability to generate new knowledge.

Meter reiterated that they "did not see evidence for significant bottlenecks in algorithmic progress, training compute, experiment compute, or data that will take effect in the next 12 months."

The Evaluation Process and Limitations

Meter deployed 2.5 full-time equivalent employees to Anthropic starting in mid-August. Their objectives were to determine: 1. If Anthropic's internal capabilities growth rate had accelerated. 2. What share of any acceleration was attributable to AI. 3. If there were important bottlenecks that could slow growth.

Meter had the ability to request any relevant information, but Anthropic retained unilateral approval/rejection power. This led to negotiations and compromises on data access. This project was explicitly not the "employee-like access" program Anthropic had previously announced.

Notably, Meter was initially slated to verify Anthropic's compute allocation (between R&D, inference, safety, etc.), but Anthropic revoked this access midway through the project, stating a desire to keep the engagement focused on risk assessment for the time being.

Unreleased Models and Benchmark Saturation

The appendices of the report revealed further insights: * New Internal Models: Anthropic is using unreleased internal models (e.g., Model 3, Model 4, beyond Mythos 5) to accelerate their own development. These models suggest a permanent, compounding slope change in capability rather than a one-time jump. * Benchmark Difficulty: Anthropic has had to make its internal benchmarks more difficult because new models are saturating them, meaning they are becoming "too good" for existing tests. * Autonomous Experimentation: Researchers inside Anthropic are reporting that, for the first time, these models are becoming genuinely useful for autonomous experimentation.

In conclusion, the Meter report provides independent, third-party validation that AI is indeed accelerating its own research and development at Anthropic. This recursive self-improvement is already underway, with no clear bottlenecks identified, suggesting a potentially rapid and compounding trajectory for AI capabilities.

  Takeaways

  • Meter's independent assessment found conclusive evidence that Anthropic's AI models are improving themselves at roughly twice the previous rate, indicating recursive self‑improvement.
  • The acceleration is attributed primarily to AI‑driven progress rather than additional hiring, compute scaling, or human researcher effort.
  • An inflection point around late 2025, coinciding with the release of Opus 4.5, marked a sharp increase in capability growth, with about 80 % of the recent acceleration linked to AI.
  • Employees reported productivity gains of more than 2× and described AI as the dominant factor behind the speed‑up, while no clear bottlenecks in compute, data, or algorithms were identified.
  • Internal unreleased models are already saturating benchmarks and being used for autonomous experimentation, suggesting a compounding feedback loop that could lead toward an intelligence explosion.

Frequently Asked Questions

Why does the Meter report attribute 80% of Anthropic's recent acceleration to AI rather than compute or hiring?

The Meter report attributes 80% of the acceleration to AI because its analysis isolated AI‑driven improvements as the primary cause, finding that model capability gains outpaced any increase in compute power, staff headcount, or human research effort.

What is the significance of the late‑2025 inflection point linked to Opus 4.5 in the report?

The late‑2025 inflection point, tied to the launch of Opus 4.5, marks the moment when model performance on code‑writing tasks surged, triggering a rapid, AI‑driven acceleration that the report quantifies as the start of a compounding growth curve.

Who is Matthew Berman on YouTube?

Matthew Berman is a YouTube channel that publishes videos on a range of topics. Browse more summaries from this channel below.

Does this page include the full transcript of the video?

Yes, the full transcript for this video is available on this page. Click 'Show transcript' in the sidebar to read it.

Helpful resources related to this video

If you want to practice or explore the concepts discussed in the video, these commonly used tools may help.

Links may be affiliate links. We only include resources that are genuinely relevant to the topic.

Full transcript is not shown on this page

This page focuses on the summary and original notes. For full verification, refer to the original YouTube video.

PDF