Anthropic AI Research Accelerates via Recursive Self-Improvement
Independent investigation reveals Anthropic's AI is building its own successors twice as fast.
1Independent Audit Findings
The independent research group METR spent 11 weeks inside Anthropic to assess AI development speeds. Their investigation concluded that Anthropic's AI is improving twice as fast as it was one year ago. This acceleration is primarily attributed to the AI models themselves performing significant portions of the research and coding work.
2Recursive Self-Improvement Mechanism
Recursive self-improvement occurs when AI systems contribute to the development of their own successors. METR estimates that approximately 80% of the current speed-up stems from this process. As models become more capable, they automate more research tasks, creating a feedback loop that continuously reduces the time required for new iterations.
3Quantifying Research Progress
Anthropic previously defined a critical threshold as achieving two years of research progress within a single year. By late 2026, the company confirmed it had crossed this line. This progress is evidenced by AI models solving complex mathematical problems that had remained unsolved by human mathematicians for the past 87 years.
4Evidence of Acceleration
METR analyzed internal test scores and training data to verify the speed-up. They observed that the performance trajectory for internal models steepened significantly around October 2025. This timing correlates with the period when Claude gained the ability to write code independently for extended durations, marking a shift in research methodology.
5Redaction and Transparency
Anthropic redacted significant portions of the METR report before public release, limiting external visibility into the specific evidence gathered. While METR maintains that the redacted data supports their findings, the lack of full transparency forces the public to rely heavily on the researchers' judgment rather than verifiable, raw internal datasets.
6Skepticism and Counterarguments
Philosopher Toby Ord suggests the observed speed-up might be partially explained by one-off jumps from larger training runs rather than pure recursive improvement. However, other experts like Buck Shlegeris argue that the redacted evidence, which remains unseen by the public, was the primary driver for METR's conclusion regarding the rapid acceleration.
7No Observed Bottlenecks
METR investigated potential constraints such as computing power, data availability, and algorithmic innovation. They found no immediate signs of these factors slowing down the development process. Consequently, the report warns that the rate of improvement could continue to increase, potentially leading to even faster cycles of model development in the future.
8Comparison to Forecasts
The current acceleration aligns with historical 'AI 2027' prediction scenarios, which previously estimated such milestones would occur later. If METR's findings are accurate, the industry may be operating several months ahead of earlier projections. This shift has led some researchers to update their probability estimates for the emergence of superintelligence.
9Regulatory and Policy Implications
Legislators like Alex Bores are now calling for federal intervention to regulate AI research automation. Proposals include banning companies from fully automating their own research processes to maintain human oversight. Such measures aim to prevent a scenario where AI systems become uncontrollable before safety protocols can be effectively implemented and verified.
10The Competitive Dilemma
AI companies face a prisoner's dilemma where stopping research could allow competitors to gain a dangerous lead. This competitive pressure discourages voluntary pauses, even as the risks of rapid, unmonitored development increase. Experts argue that without mandatory, independent oversight across all frontier labs, the current trajectory of AI development remains largely unchecked.
11Limitations of External Scrutiny
METR's investigation was conducted under terms set by Anthropic, which controlled access and publication rights. This dependency creates a conflict of interest, as researchers may fear that overly critical reports could jeopardize future access. Experts emphasize the need for standardized, mandatory reporting requirements for all companies developing frontier-level artificial intelligence models.
12Future Outlook and Risk
The transition to AI-driven research represents a fundamental shift in technological development. Whether the speed-up is 1.4 or 2.5 times faster, the trend toward recursive improvement is clear. The primary challenge remains ensuring that safety and control mechanisms can keep pace with the rapidly evolving capabilities of these increasingly autonomous systems.
Original article: 80,000 Hours — Anthropic AI Research Accelerates via Recursive Self-Improvement