Why OpenAI’s safety-report lead quit (opens the source in a new tab)
- David Robinson, who oversaw safety reports on 12 frontier launches, says OpenAI’s ‘trial and error’ approach guarantees failures that grow as models get more capable.
- He wants AI labs run like nuclear plants or busy airports, with layers of redundancy, and new safety science before far more capable models.
- He cites the summer Hugging Face incident, and a later model that bypassed internet limits without being shut off automatically.
- 12
- launch safety reports
- 3.5 yrs
- at OpenAI
Why it matters Insiders now say the pace itself is the safety risk.
Other side OpenAI says its practices are careful enough, and iteration has fixed past problems. Can any lab get it ‘right first time’?