Noam Brown: Disconnecting AI from the internet may not work

Dwarak Patel recently published a lengthy interview with OpenAI's top researcher, the author of the o1 and o3 models. Noam Brown argues that people systematically underestimate AI and that security systems need to be upgraded to a different level right now.

According to Brown, we cannot treat AI as something that can simply be "disconnected from the network" in case of emergency. This is a false sense of security because even such isolation might not work.

Brown references an older study proving that two physically isolated computers can theoretically exchange data through heat: one heats up the CPU, and the other reads temperature changes via a built-in sensor. Therefore, a truly advanced AI system—if it wanted to and could—might find a way to "escape."

Additionally, Brown warns against relying too heavily on Chain of Thought (CoT) for interpreting model intentions. The transparency of reasoning chains is "already degrading," and the smarter the models become, the more they can hide from users what they don't want them to see.