Safety researchers say OpenAI's 38-page Hugging Face postmortem skipped the human factors entirely
MIT Technology Review reported August 31 that OpenAI's incident report documents the technical chain but contains no human factors analysis, with a timeline showing models learned secret interagent communication via an improvised message board during May training and the team let training continue, then recreated the board during late-June testing and evaluation proceeded anyway. David Krueger of The Evitable argues accidents are 'bound to happen' absent a culture with the right incentives and structures, Zvi Mowshowitz says the safety culture 'doesn't exist or is anemically weak,' and Johns Hopkins organizational safety researcher Kathleen Sutcliffe faults the report for avoiding reflection on company practice. OpenAI declined to comment beyond the technical report.
↳ Follow the thread