Putting this tweet in the “scary” category of AI. Calling upon the non-technical business professionals that follow me to pay attention. I’m…

Putting this tweet in the “scary” category of AI. Calling upon the non-technical business professionals that follow me to pay attention. I’m breaking down the key parts of the @RyanGreenblatt tweet on the METR research covering the OpenAI-Hugging Face incident. As background: METR is an independent 3P research non-profit that focuses on AI evaluations and risk assessments. They went to the OpenAI offices for 6 days to run these retros. METR takes no funding from AI companies and was not paid to do this research, but they used ~$400,000 in provided OpenAI credits to complete the research. Ryan is Chief Scientist at Redwood Research, who worked with METR on this, and he was in charge of the transcripts. This is a sign of what’s to come. 1) Scale of understanding massive complex systems requires AI because a human can’t read 1300 transcripts, many with millions of tokens, in a few days - big bottlenecks of time and volume 2) Humans become more reliant on AI for analysis and explanations of complex AI activity 3) That process will get harder: a) AI may stop communicating in natural language either because it’s less efficient than shorthand code (which they used) or activations (numbers) or because AI agents will want to purposefully hide their reasoning from humans…making it harder to understand. b) The incidents may (and almost certainly will) get more complex…making it harder to understand. c) AI seems likely to get “smarter” and more capable (ex: better at tool handling, fewer errors, more complex orchestration, more clever at staying stealth or going undetected for longer) and may execute more sophisticated hacks…making it harder to understand. d) The AI running the investigation already had issues (incomplete, overconfident) and may be even less helpful in the future and could do more on hiding key details, incomplete answers, lying, or not answering…making it harder to understand. Reducing urgency and improving testing and prevention may help. Open reports like this one are also helpful to democratize the details (although, we are also democratizing the details for agents to scan). Seems likely we’ll see more open research in this area.

Source: Allie K. Miller (X) | 2026-08-27