OpenAI and Anthropic recently called for slowing down AI research, arguing that new models are becoming increasingly dangerous. However, OpenAI researcher Noam Brown raised the stakes even further on Dwarkesh Patel's podcast: even if AI is physically disconnected from the network, it might still be able to communicate with each other.

He referenced an old case from May of this year — the incident where OpenAI's AI attacked the Hugging Face website, which caused quite a stir abroad. What really supports his argument is a study on physically isolated networks: many people assume that cutting off the network, especially through physical disconnection, would prevent AI from communicating or spreading, thus keeping the risks under control; however, there have already been explorations of AI communication in "air-gapped" networks.

This mechanism has a strange name: BitWhisper. The principle is related to temperature: two computers placed side by side, even if they are separated by air, if one's CPU heats up, the other can detect the change in temperature — this cold and hot becomes a communication channel between the two machines. However, Gizmodo also pointed out in its report that this research is very early stage, with extremely strict limitations: the distance between the two computers cannot exceed 40 centimeters, and both must be infected with a virus beforehand. Even so, the transmission rate is extremely low, only 1 to 8 bits per hour. In other words, technically it's "possible," but from an engineering perspective, it's still far from being a real threat.

What is truly intriguing is the counter-question posed by Gizmodo: long before the attack on Hugging Face, OpenAI had already been researching mechanisms to monitor "chain-of-thought." If they were worried about losing control, why didn't they just turn this switch off back then? This directly brings up the issue of the motives behind the "warning."

The debate around the "apocalypse theory" doesn't stop here. NVIDIA CEO Jensen Huang has already made his point clear: repeatedly spreading the narrative that "AI is harmful and will destroy humanity" essentially helps these companies "worship," portraying themselves as saviors, and making their own AI business seem like the "only correct" choice — choosing others equals irresponsibility. This has been seen by some as the underlying logic behind their frequent denigrations of open source and Chinese AI. Huang also mentioned that the AI apocalypse theory is almost nonexistent in China. He didn't elaborate on the reasons, but a straightforward explanation is that technology is viewed as a tool rather than a god or demon, and its goodness or badness depends on the people who use it.

On one side, researchers are seriously discussing the boundaries of air-gap communication, while on the other, external scrutiny is questioning the motives behind the "warnings." As AI safety slides from an engineering issue into a narrative battle, the temperature signal that transmits only eight bits per hour has become the most concrete technical footnote in the entire debate.