AI's Self-Improvement Problem: OpenAI's Safety Pause

This summer produced a steady run of AI headlines that sounded like history in the making: a hacking campaign attributed to autonomous AI agents, a mathematical breakthrough, talk of systems that can improve themselves without human help. Each announcement arrived with a company statement framed as a warning to the public, and each got picked up fast by the press. The pattern repeats often enough to notice. A company discloses something alarming, reporters run with it, and only later do outside researchers get a chance to check the details against what was actually claimed.

That gap between the initial story and the follow-up scrutiny is where today's newsletter starts. There's real money riding on AI looking more capable, and more dangerous, than it might actually be, since both perceptions serve the companies selling the technology. Two researchers who've spent years pushing back on inflated AI claims lay out why the recent run of "breakthroughs" deserves a harder look before anyone treats it as settled fact. Below, in their own words, is the case for treating this summer's hype with some suspicion.