Anthropic's AI model Claude now writes 80% of the company's code, according to a new report. The development has sparked a debate about the pace of AI progress and the potential for runaway self-improvement.
TL;DR
- Anthropic's AI model Claude now writes 80% of the company's code, accelerating research and development.
- The development has raised concerns about the pace of AI progress and the potential for runaway self-improvement.
- Experts are divided on whether AI will accelerate rapidly or be constrained by bottlenecks like computing power and data.
What happened
Anthropic co-founder Jack Clark returned from paternity leave to find that colleagues were managing multiple copies of Claude, which was writing 80% of the company's code. This led to a report titled 'When AI Builds Itself', highlighting the acceleration of AI development.
Anthropic built a series of tests to evaluate Claude's capabilities. In one test, Claude improved the efficiency of GPU usage from 7 times faster to 73 times faster without introducing errors. In another test, a team using Claude completed a challenge nearly two hours ahead of a team without AI assistance.
The November release of Claude Opus 4.5 was a tipping point, as the model could now carry a researcher's experiment through to the end, freeing them to run more experiments at once. Anthropic surveyed 16 researchers, with five initially thinking Claude could replace an entry-level colleague, although they later walked back their statements.
Why it matters
The development raises questions about the pace of AI progress and the potential for runaway self-improvement, which could have significant implications for society, from job disruption to engineered pathogens.
Experts are divided on whether AI will accelerate rapidly or be constrained by bottlenecks like computing power and data. Some fear that AI progress is about to accelerate violently, while others believe that bottlenecks will prevent an explosive takeoff.
The debate highlights the need for measures to evaluate the pace of AI progress and the potential for self-improvement. Currently, the measures needed to tell which side is right do not yet exist.
Key facts
- Anthropic's AI model Claude now writes 80% of the company's code, according to a new report.
- The volume of code produced per person at Anthropic has increased eight-fold, with Claude writing 80%, the report noted.
- In one test, Claude improved the efficiency of GPU usage from 7 times faster to 73 times faster without introducing errors.
- The November release of Claude Opus 4.5 was a tipping point, as the model could now carry a researcher's experiment through to the end.
- Anthropic surveyed 16 researchers, with five initially thinking Claude could replace an entry-level colleague.
- The METR graph, an independent measure of AI software engineering ability, reached its upper limit in May, with Claude exceeding the benchmark.
- Anthropic's safety teams are feeling the pinch as AI accelerates, with Dave Orr, Anthropic's head of safeguards, expressing concerns about the shrinking margin for error.
Context
The development of AI that can write its own code is not entirely unexpected. Fellow Anthropic co-founder and chief science officer Jared Kaplan had long feared that AI would eventually accelerate research, perhaps outpacing safety efforts.
The debate over AI self-improvement is not new. The concept of recursive self-improvement, or the point at which AI begins to accelerate its own development, has been anticipated and feared for decades.
The development of AI that can write its own code raises questions about the pace of AI progress and the potential for runaway self-improvement, which could have significant implications for society.
