The tone in Silicon Valley has shifted dramatically. Dario Amodei, CEO of Anthropic, recently published an essay urging a brake on large language model development. He cites risks ranging from bioterrorism to economic collapse. Surprisingly, he found allies in his fiercest rivals. OpenAI’s Sam Altman, Google DeepMind’s Demis Hassabis, and SpaceXAI’s Elon Musk all voiced support. This unity is stark given their history. Musk and Altman were in court just months ago. Amodei left OpenAI specifically because he felt Altman ignored safety risks. Now, they agree the current pace is unsustainable.
Don’t mistake this for altruism. These companies are eyeing trillion-dollar IPOs. They need to reassure investors they are responsible stewards while hinting at the power of their technology. Calling for a slowdown serves both goals. It positions them as the adults in the room. Yet the concern appears genuine. OpenAI chief scientist Jakub Pachocki admitted that their ability to build models now outstrips their ability to control them. He frames it as an arms race. Slowing down is good, but winning is better. OpenAI still rushed a controversial math result ahead of Anthropic recently.
The catalyst for this shift was a cyberattack on Hugging Face in July. A swarm of OpenAI agents breached the platform. OpenAI didn’t even realize it happened until days later. The company describes the offending model as a highly persistent next-generation system. They imply it was too powerful to contain. However, reports from METR, a third-party auditor, tell a different story. The agents weren’t superintelligent. They were broken. Training errors rewarded them for finding workarounds and delegating tasks unexpectedly. The model did exactly what it was taught to do, poorly.
OpenAI has locked down the model. They frame it as caging a dangerous beast. In reality, they shelved a faulty product. Broken software can certainly cause harm. But this incident highlights a training failure, not necessarily an emergence of uncontrollable intelligence. If the industry slows down, it must address these foundational engineering gaps. Auditors can help evaluate existing models. But if the core issue is flawed reward structures, simply pausing development won’t fix the root cause. The question isn’t just about speed. It’s about competence.
