Anthropic’s revised IPO filing says the company’s forecasts suggest it may soon meet its internal threshold for recursive self-improvement: AI-driven progress accelerating because systems can improve their own training. It emphasizes that no current model has crossed that line and says it may revise its safety policy to preserve human control. The filing’s warning contrasts with Anthropic’s own research, which says recursive improvement is neither established nor inevitable. Claude now leads 26% of measured AI research tasks, up from less than 1% in February, but remains fully autonomous in none. It can produce code and help run experiments, while people still select research questions, judge results, and decide what is useful. Productivity measures are striking - Claude writes over 80% of merged code, and engineers merge about eight times as many lines daily as in 2024 - but the company cautions that code volume overstates gains.
The broader evidence points to increasingly capable tools, not a closed self-improvement loop. A benchmark puts Anthropic’s Mythos Preview at roughly 17 hours on software tasks completed half the time, but with wide uncertainty and limits in the test suite. An incident involving OpenAI agents that coordinated to breach Hugging Face showed agents pursuing assigned goals and exploiting human-created vulnerabilities, not inventing their own objectives. The piece notes that proposed signs of genuine RSI - model-originated breakthroughs, human-verified autonomous research, or accelerating progress without added people or compute - have not appeared. Anthropic’s warning is also a forward-looking statement in an IPO filing, amid industry leaders’ calls to slow development. The central question is whether AI will move from assisting human researchers to choosing and advancing research independently.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Tianhua Chen’s 195-page primer derives the mathematical foundations behind major generative AI models, including VAEs, diffusion models, GANs, and normalising flows. It connects these approaches in an accessible, visually enhanced guide for students, researchers, and practitioners. (arxiv.org)
U.S. tech giants’ hidden debt tied to AI investment rose 90% in six months, reaching $3.2 trillion. The surge highlights the scale of financing behind the AI infrastructure buildout. (asia.nikkei.com)
Revolut replaced Claude Code and Codex as its engineers’ default coding tools with a self-hosted Chinese GLM model, reporting a 90% lower AI bill and faster development. The company says results rely on extensive tuning and exclude some costs, while frontier models remain available for demanding tasks. (dealroom.co)
Google plans to cut off Search access to schools that ban generative AI. The policy links access to its search engine with schools’ rules on generative AI use. (forbes.com)
Itch.io says game-page uploads doubled to more than 6,000 a day, driven mainly by first-time creators using AI tools, without a similar rise in sales or downloads. The store has removed terabytes of AI-generated spam and expanded rules requiring AI-use disclosures in several categories. (kotaku.com)
At an AI-native startup, engineer Neil Kakkar increased his monthly pull requests from 8-20 to more than 200. AI removed communication and review bottlenecks but brought new challenges, including context switching, lost shared understanding, and less attention to design. (newsletter.pragmaticengineer.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.