#recursive self-improvement
5 articles tagged recursive self-improvement
Tagged articles
Articles tagged recursive self-improvement, newest first.
AI NewsMETR Audit Finds Anthropic AI R&D Running at Twice Its Historical Pace
METR: Anthropic improves models on well-specified tasks at about 2x its past pace. About 80% of the speedup is attributed to AI tools, not compute growth or hiring. Capability index growth rose from 15 points a year to over 29 after Opus 4.5. No bottleneck found for the next 12 months, but AI R&D is not yet fully autonomous. Weigh the caveats: redactions, a 1.3x to 3x range, and a narrowed audit scope.
AI NewsAnthropic S-1 RSI Disclosure: 2x Research Speedup, Not a Runaway Loop
Anthropic's S-1 says it may soon meet its own recursive self-improvement threshold. That threshold is a 2x research speedup, roughly an acceleration factor of 0.5. A runaway loop needs R of 1, and compute limits plus safety checks still block it. Reportedly only about 1 percent of Fortune 500 staff max out enterprise AI seats. Redesign end-to-end workflows and train staff instead of just buying licenses.
AI NewsMETR Report: AI Progress Doubles in Speed, Raising 2027 Self-Improvement Debate
METR report puts AI capability growth at about 2x the pace of past trends. Strongest evidence redacted because Anthropic barred public release of the data. Palisade Research's director moved his autonomous AI research estimate toward 2027. An estimated 4.5x speedup is algorithmic, so tracking hardware alone falls short. Judge models on long end-to-end tasks rather than synthetic benchmark scores.
AI EthicsDemis Hassabis on Frontier AI Governance: Keep Humans in the Loop
Hassabis calls self-improving AI with no human review the riskiest path to AGI. Safeguards he wants: hardened sandboxes, runtime monitoring, cross-lab red-teaming. He backs an industry self-regulatory body with government oversight and audits. He sees no evidence of consciousness in today's models, though experts differ. He expects 6 to 8 frontier labs to lead and urges nations to invest downstream.
AI EthicsAnthropic's S-1 Calls AI Building AI Both a Safety Risk and Its Core Strategy
Anthropic's S-1 calls AI building stronger AI a major safety risk and a core edge. Claude led roughly 25% of Anthropic's internal tasks by August 2026, from near zero. Filing admits AI acceleration shows up only in lagging indicators, after the fact. Epoch AI benchmark puts Claude Fable 5 at about 2% of human original invention. Read S-1 risk sections carefully: they warn investors but also market the company.
Get new articles by RSS or podcast
Add the RSS URL to your RSS reader or the podcast URL to your podcast app to get every new article as soon as it is published. The podcast reads out the headline and key points.
How to subscribe