Reading up on METR

Ajeya Cotra forecasts AI self-sufficiency within 10 years; Timothy B. Lee gives a 50-year median, debating robotics, tacit knowledge, and profit incentives.

Anthropic presents internal data showing AI systems now write 80% of its code and accelerate research, trending toward autonomous recursive self-improvement.

Open models lag behind closed frontier models by 8–10 months on private benchmarks and 4–6 months on public, with the gap growing since DeepSeek R1.

A chart from the nonprofit METR has become an industrywide obsession for measuring the rapid development of large AI systems.

LLMs do not provide a revolutionary productivity gain in software development because the essential difficulty lies in specification, design, and testing, not code generation speed.

AI has entered a new agentic era where models can autonomously complete hours of human work, and exponential capability improvements are now driving radical organizational experiments like StrongDM's "Software Factory" that ships code without human touching it.

AGI capable of most cognitive work could arrive by 2028–2034, but deployment will lag capability due to verification bottlenecks, uneven automation, and institutional friction.

METR’s late 2025 experiment on AI developer productivity suffers from selection bias because many developers refuse to work without AI, yielding unreliable speedup estimates.

Opus 4.6 delivers 14 times more useful work per compute-minute than Codex 5.3, a metric called Intelligence Yield derived from METR Time Horizons data.

OpenAI's Codex successfully built a design tool from scratch over 25 hours, demonstrating long-horizon agentic coding with milestone-based planning and continuous verification.

Software engineering accounts for half of AI agent usage, leaving healthcare, legal, and other verticals as greenfield opportunities for founders to build domain-specific agents.

Matt Shumer argues that frontier AI models released in early 2026 can autonomously build software and are on track to eliminate most white-collar knowledge work within a few years.

Claude Code's ease of building apps exposes that most people lack 'software vision' to identify software-shaped problems, and the tool can even decrease productivity by enabling procrastination on non-software tasks.

Long-horizon agents that can figure things out autonomously are functionally AGI, and they will become widespread in 2026.

AI progress follows compute scaling, making evaluation increasingly difficult but promising radical improvements in infrastructure and economic decision-making.

Anthropic's internal survey and interviews reveal that engineers using Claude report 50% productivity gains, a 2-3x increase from last year, while raising concerns about skill atrophy and reduced collaboration.

Coding agents like Codex can now sustain multi-hour reasoning tasks, transforming the entire software development lifecycle by delegating mechanical work across planning, design, build, test, review, and deployment.