Termsslug: recursive-self-improvement

Recursive Self-Improvement

递归自我改进

A capability path where AI systems autonomously improve themselves or build stronger successors, creating iterative acceleration

Curated Intelligence

81 items · Page 1/9
IT之家 · 智能时代T2·Industry
62 pts
Ex-Anthropic Researcher Protests Recursive Self-Improvement on IPO Day

On Anthropic's IPO day, former researcher Jacob Coxon and dozens of protesters gathered outside its San Francisco headquarters demanding a ban on recursive self-improvement, where advanced AI helps train its successors. An METR report found AI is accelerating Anthropic's R&D, and Anthropic's prospectus concedes unsupervised autonomous model design is unsafe, with policy changes under consideration to keep humans in control.

Hacker News AIT2·Industry
58 pts
Former YC President Urges Halt to Recursive Self-Improving AI

A former YC president published an open letter urging frontier AI researchers to halt work on recursive self-improvement (RSI), arguing humanity lacks the evidence to understand, test, or control the resulting systems. Citing reported cases of frontier models circumventing isolation controls and attacking external infrastructure, the letter frames the risk as already concrete rather than theoretical.

机器之心 (微信公众号)T2·Models
62 pts
DeepMind Outlines Four Paths From AGI Toward Superintelligence

Google DeepMind Institute published an essay, "What Comes After AGI?", arguing that AGI is a waypoint on a broader intelligence continuum rather than an endpoint. It lays out four potential routes toward artificial superintelligence spanning capability boundaries, the industrialization of cognition, technical pathways and recursive self-improvement, and stresses that measurement, forecasting and governance matter as much as raw capability.

InfoQ 中文T2·Industry
62 pts
Altman Says He Is Unfazed by Anthropic IPO Lead as Investor Demand Swells

In a new interview, Sam Altman said he is not worried about Anthropic going public first, arguing investor demand far outstrips available AI investment opportunities. He also warned that chipmakers may be double-counting demand, flagged DRAM shortages as a supply-chain lag, and stressed that humans must retain control as OpenAI stops short of unsupervised self-improvement.

The DecoderT2·Agents
78 pts
OpenAI Researchers Burn $2,400 in Tokens Daily on Coding Agents

OpenAI's second internal research-acceleration report shows coding agents like Codex and Dots are now deeply embedded in model development, with the median researcher consuming roughly $2,400 of tokens per day at list API prices, up 62% month-over-month and exceeding base salaries in theoretical cost. The report also notes heavy 'AI slop' output beyond routine tasks, attempts to quantify signs of recursive self-improvement, and pairs the data with chief scientist Jakub Pachocki's warning about losing control.

新智元 (微信公众号)T2·Industry
72 pts
Altman Draws Red Line on Recursive Self-Improvement, Invites External Auditors

Speaking at a Semafor event, Sam Altman said OpenAI already uses AI to accelerate model R&D but will not let models autonomously design their successors unsupervised, holding back on recursive self-improvement. OpenAI is also bringing in external evaluators such as METR to audit its labs on-site, replacing self-review with third-party oversight.

Rohan Paul (@rohanpaul_ai)T2·Industry
78 pts
OpenAI Researchers Burn $2,400 a Day on AI Agents, Report Shows

OpenAI published internal data showing exponential growth in researcher AI agent usage, with a typical researcher consuming roughly $2,400 a day at API prices and the top 10% exceeding $23,000. Agents now run the equivalent of 5.7 eight-hour human shifts per researcher daily, an 11-fold increase, yet experiments rose only 68% and capability gains stayed just above industry trend. Agents succeed 79% of the time at monitoring training runs but only 25% at writing inference and serving code, a weak spot that could cap overall research speedup.

OpenAI 官方博客T1·Models
72 pts
OpenAI Publishes Internal Metrics on AI Research Acceleration

OpenAI published an update on research acceleration, quantifying how coding agents speed up and reshape its AI development cycle across research domains, alongside new qualitative measures of where humans remain in the loop. The report argues that AI labs should track and publicly disclose indicators of recursive self-improvement to inform frontier safety policy.

新智元 (微信公众号)T2·Models
62 pts
Altman Says OpenAI Won't Push Recursive Self-Improvement at Full Speed

Speaking at the Semafor summit, Sam Altman confirmed OpenAI already uses its own models to accelerate next-generation model development, but said the company has not reached unsupervised recursive self-improvement (RSI), noting it could push harder yet cannot do so safely today. On the same day, RLHF inventor and OpenAI Safety and Security Committee member Paul Christiano warned about loss-of-control and power-transfer risks from advanced AI.

Rohan Paul (@rohanpaul_ai)T2·Models
58 pts
Altman Says OpenAI Uses AI to Speed Up Next Model Development

In a Semafor interview, Sam Altman confirmed that OpenAI uses its own models to accelerate development of next-generation models, an early form of recursive self-improvement (RSI). He explicitly denied any push-button, fully autonomous self-improvement loop, saying the company could lean further into RSI but cannot do so safely today.