Skip to main content

Anthropic Accelerates AI Research as METR Warns of RSI Risks

An 11-week safety audit by METR reveals Anthropic is doubling its R&D pace using internal AI models, prompting debate over recursive self-improvement and report redactions.

S
Written byShtef
Read Time4 minutes read
Posted on
Share
Anthropic Accelerates AI Research as METR Warns of RSI Risks

Anthropic Accelerates AI Research as METR Warns of RSI Risks

Inside the Safety Audit Revealing Rapid Progress Toward Self-Building Models

Anthropic is leveraging advanced internal AI models to build more capable successors, doubling its pace of research and development over the past year. A landmark 11-week safety audit conducted by research organization METR reveals that frontier models are now actively accelerating their own engineering pipeline. The report has triggered alarm across the AI community, intensifying debates over recursive self-improvement and transparency as Anthropic nears its public debut.

Key Details

The investigation by Model Evaluation and Threat Research (METR) examined Anthropic's internal engineering workflows and model training cycles. According to the published evaluation, Anthropic's models are advancing twice as fast as they did twelve months ago. Crucially, benchmark metrics show the lab accomplished approximately two years of historical deep learning progress within the last year alone.

Key findings documented by the audit include:

  • AI-Driven Acceleration: METR researchers concluded that the speedup is largely attributable to internal AI tooling rather than raw compute scaling, marking a fundamental shift from hardware-driven gains to algorithmic self-assistance.
  • Enormous Productivity Multipliers: Internal surveys indicated that engineers utilizing Anthropic's unreleased frontier models reported productivity multipliers averaging twelve times baseline performance, doubling previous self-reported gains from the Mythos 5 testing phase.
  • Absence of Diminishing Returns: Evaluators identified no observable bottlenecks or scaling ceilings currently constraining this acceleration, suggesting research velocity could climb further.
  • Controversial Heavy Redactions: Anthropic exercised contractual rights to heavily redact evidence from the public ~100-page dossier, citing competitive risks, which drew sharp criticism from independent safety evaluators.

What This Means

The METR audit provides empirical evidence for a long-theorized milestone: artificial intelligence accelerating the creation of next-generation artificial intelligence. While frontier labs have previously used language models to generate synthetic datasets or automate routine pull requests, Anthropic's workflow represents systematic integration of unreleased models into frontier research and architectural iteration.

The report also lays bare the profound tension between safety advocacy and commercial competitiveness. While Dario Amodei has publicly urged the industry and policymakers to decelerate frontier training runs, Anthropic's internal teams are utilizing proprietary models to compress research timelines. Citing rival risks to justify heavy redactions undermines external oversight at the exact moment independent verification is most essential.

Technical Breakdown

METR's findings shed light on the mechanics powering this research acceleration, as well as the immediate limits observed during the evaluation:

  • Assisted vs Autonomous Workflows: The documented speedup stems predominantly from human-in-the-loop acceleration, where senior researchers leverage unreleased frontier systems for rapid code refactoring, mathematical derivation, and hypothesis generation.
  • Emerging Autonomous Loops: Evaluators noted nascent signs of autonomous contributions beginning to shoulder end-to-end tasks, raising the prospect that human oversight could become a latency bottleneck.
  • Compute Allocation Divergence: Anthropic cancelled a planned assessment phase that would have tracked compute allocation between capability scaling and alignment verification, deferring resource scrutiny to future engagements.

Industry Impact

The revelation that recursive self-improvement dynamics are already taking root changes the strategic calculus for competitors and enterprise software leaders. If internal productivity gains compound at 12x, the gap between well-capitalized frontier labs and open-weight ecosystems could widen dramatically.

The report has provoked immediate regulatory fallout. Following revelations in the audit, New York State Assemblyman Alex Bores called on Congress to consider legislative bans on unconstrained recursive self-improvement before year's end. With political shifts following the midterm elections and scrutiny mounting over data center expansion, labs face intensified pressure to demonstrate that self-accelerating research remains under strict human control.

Looking Ahead

Anthropic co-founder Sam McCandlish indicated the company plans to release its own technical report on the velocity of AI progress alongside strengthened safety commitments in the coming weeks. However, independent researchers argue that voluntary disclosures are insufficient when models are building their successors behind closed doors.

As frontier labs approach recursive thresholds where AI systems independently architect neural architectures, the frontier AI safety regime will face its greatest test. The critical question is no longer whether models can accelerate their own development, but whether governance frameworks can keep up before human supervision becomes obsolete.


Source: The Verge(opens in a new tab) Published on ShtefAI blog by Shtef ⚡

Recommended

Related Posts

Expand your knowledge with these hand-picked posts.

OpenAI Quarterly Revenue Crosses $10B Amid Enterprise AI Surge
AI News

OpenAI Quarterly Revenue Crosses $10B Amid Enterprise AI Surge

OpenAI records historic $10 billion in quarterly revenue as surging enterprise demand and agentic AI models drive annual run rate toward $50 billion.

New York Launches RAISE Act Oversight for Frontier AI Labs
AI News

New York Launches RAISE Act Oversight for Frontier AI Labs

New York State opens its registration portal under the RAISE Act, requiring major frontier AI developers like OpenAI and Anthropic to report incidents and publish safety frameworks.

DeepMind AlphaProof Code Formally Verifies Signal Protocol
AI News

DeepMind AlphaProof Code Formally Verifies Signal Protocol

Google DeepMind introduces AlphaProof Code, an agentic framework that generated six million lines of Lean proofs to mathematically verify the core Signal cryptographic protocol.