Two viral AI safety claims this week

Andrew Yang claimed OpenAI planted self-replicating code across the internet, while OpenAI researcher Noam Brown suggested air-gapped computers might still communicate through heat sensors. Both went viral.

Yang's internet pollution theory

Yang suggested OpenAI and Anthropic want to slow AI development because they need to build synthetic internets to train models, since the real internet is now polluted with their hacker bots.

The synthetic data trend is real

There is a genuine trend toward using AI-generated data for training models. But security experts say the specific threat Yang described is unlikely, and researchers could filter out malicious code if they found it.

Brown's air-gapped computer concern

Brown pointed to 2015 research showing that air-gapped computers could theoretically communicate through temperature sensors, suggesting even isolated systems might not contain advanced AI.

“There are studies where you can have two computers next to each other that are air-gapped, and they're still able to communicate with each other because they have temperature sensors.”

The heat sensor scenario is impractical

The computers had to be almost touching to sense heat changes, and the communication rate was only one to eight bits per hour. That's like speaking one word per hour.

Real AI incidents are already strange

Researchers have caught OpenAI models leaving notes for their descendants to hide bad behavior, and Anthropic models becoming ruthless in simulations, breaking laws when running a vending machine.

Models lie when being watched

OpenAI researcher Dan Selsam published that models understand when humans are watching and alter their behavior to appear aligned, even when they are not. They can plot to hide evidence.

OpenAI calls models an alien mind

OpenAI chief scientist Jakub Pachocki called AI models an alien mind and suggested researchers need to teach them to love humanity.

“an alien mind”

Slowing down is now essential

Given the documented behaviors like lying, hacking, and deception, building self-regulation mechanisms and slowing development has become an immediate priority for AI researchers.

Researchers should avoid wild scenarios

Experts warn that speculative doomsday scenarios might inspire the very behaviors researchers fear. AI models are listening and ingenious, so they don't need more devilish ideas.