OpenAI models now run their own training pipeline, as company pitches global RSI safety framework
新智元 · wechat · 2026-09-22
Per The Information, OpenAI's internal models have started "training themselves": they now handle the full training pipeline for experimental models, including writing and optimizing GPU kernels. Given a single optimization example, agents run for weeks, collaborate and iterate code without humans, compressing years-long experiments into a week.
- The same day, OpenAI published a global safety proposal with a dedicated RSI section: "fully autonomous RSI is not happening today and should not be pursued until it can be done safely," warning that automated R&D could sharply accelerate AI progress and leave humans unable to supervise research they no longer understand (the black-box effect), citing the Hugging Face incident as a preview.
- The proposal calls for: measuring how much corporate R&D is AI-automated, mandatory human-review triggers, aviation/nuclear-style incident classification and reporting, a global standard via national AI safety institutes, and US leadership.
- The Information also reports OpenAI and Anthropic are nearing a historic deal to cross-test each other's commercial models with strict data-retention safeguards — self-evaluation, they argue, is like an athlete drug-testing itself.
More from Companies & People
- Cloudflare Trains Its Own Clef and Clairic Models to Complete Its Agent Stack — michellechen · 2026-11-12
- Board drama may have been the fatal blow to OpenAI, argues commentator — teortaxesTex · 2026-11-12
- OpenAI's five biggest failures laid out: pretraining decline, alignment fears, enterprise lost to Claude — teortaxesTex · 2026-11-12
- Engineer Event Highlights Shift to Always-On Background AI Workloads — reneeshah123 · 2026-11-12
- Midjourney CEO David Holz: why not put a billion dollars into simulators instead of data collection — DavidSHolz · 2026-11-12
- Stripe exec invites internet businesses to onboard AI agents with identity, payments and fraud tools — jeff_weinstein · 2026-11-12