Artificial intelligence researchers have long imagined systems that help create more capable versions of themselves. The idea, known as recursive self-improvement, is now a major focus for AI-lab leaders, who see it as potentially decisive for staying at the technological frontier and transformative for industry, the economy and humanity. The central question is whether AI can increasingly contribute to the research and development that produces better AI - and whether that feedback loop will accelerate progress.
A chart based on Anthropic’s internal estimates tracks AI’s role in developing AI models from August 2025 through August 2026, distinguishing work with no AI involvement from minimal assistance, collaboration and AI-led research. It depicts growing participation across those levels, making the shift toward machines helping build smarter machines tangible. But the chart also draws an important boundary: as of August 2026, none of Anthropic’s model R&D was fully autonomous. AI’s expanding role therefore signals a developing process, not a completed handoff from human researchers. The significance lies in how quickly AI is moving from tool to collaborator and, in some work, leader - an evolution that could reinforce itself as stronger models become more useful in research.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
George Hotz argues that AI could bring either rigid centralized control or chaos through fabricated media, cyberattacks, and autonomous systems, with chaos being the less dangerous outcome. He sees decentralization as a possible response to vulnerable centralized infrastructure and warns that AI could upend work, politics, and daily life. (geohot.github.io)
Sam Altman argued that cloud data centers use scarce DRAM more efficiently than personal computers running local AI, where memory often sits idle. Centralizing memory could benefit OpenAI while worsening component shortages and giving consumers less control over their computing. (gizmodo.com)
An AI agent selected a gene therapy vector and brain target, then used a motorized rig to inject a mouse with epilepsy while monitoring a camera feed. A licensed veterinarian approved each irreversible step and held a physical kill switch. (twitter.com)
Proton’s AI manifesto presents AI as a tool to support software engineers rather than replace their judgment. It emphasizes keeping humans responsible for code quality, security, and privacy. (proton.me)
Gödel and Turing showed that Hilbert’s dream of a complete mathematical foundation and an algorithm to decide every theorem was impossible, but their work helped shape modern computing. AI agents paired with proof assistants like Lean are now bringing parts of Hilbert’s “proof machine” to life, with potentially major consequences for mathematics. (remy.wang)
China’s MOFCOM and MOST proposed restricting exports of 23 AI technologies, including frontier model weights, agent architectures, and offensive cyber capabilities. A pre-export registration requirement is already in effect, while the proposed restrictions remain open for public comment until November 15, 2026. (labs.cloudsecurityalliance.org)
Today's best Hacker News stories, summarized and screenshotted, one email a day.