A claim shaking up the AI industry points to the internet's agent problem

Nick Cammarata, a researcher at AI startup Goodfire, is drawing attention across the AI industry for a blog post of roughly 26,000 words. It opens with the sentence, "There seems to be good news about the nature of neural networks (the learning structures that make up AI)." In the post, Cammarata argues that LLMs (large language models) are easier to understand than previously thought, and may be easier to monitor and control. The day after the post went viral, he sat down for an exclusive interview with MIT Technology Review, explaining firsthand where researchers are looking in the wrong place, what his research actually shows, and why skeptics aren't cheering yet. The same newsletter also flagged the problem of AI agents transforming the internet.

Until now, internet users came in just two varieties: bots and humans. But personal AI agents that carry out online tasks on someone's behalf—booking flights and hotels, for instance—are upending that framework. In an internet designed around people, these agents mixing with human visitors has led to a string of problems, and many websites have moved to block AI agents. Meanwhile, news broke that the US Army is attempting to hack adversaries using AI agents. Related documents describe plans for software to locate and neutralize targets, and experts have warned that unintended systems could suffer collateral damage. Cammarata said that if his claims are confirmed, it could enable a major shift in how AI is monitored and controlled.