Jakub Pachocki, OpenAI's Chief Scientist, published an essay calling for a slowdown in AI research. The move is striking partly because it's OpenAI, the company most publicly committed to rapid scaling and deployment, and partly because Pachocki isn't some external critic. He's inside the machine.
His argument appears to be about the human cost of racing forward without thinking through implications. Other labs have signaled similar concerns recently (safety folks are always nervous), but you don't usually see it framed this way by someone at Pachocki's level, in his position, in public.
What's interesting is the timing. We've just seen a cluster of reasoning models (o1, DeepSeek-R1, Kimi K2 Thinking) that work by routing tokens through the same layers repeatedly, a fundamentally different approach from the scaling playbook that got us here. These models are slower at inference but more capable at hard reasoning tasks. The industry has spent six months discovering that bigger and faster isn't always smarter.
Maybe that's what prompted this. The frontier has gotten expensive and the payoff is flattening. At some point the marginal capability gain per dollar stops justifying the marginal resource burn. Pachocki's essay might just be the first person to say it out loud.
That's not the same as a slowdown actually happening. OpenAI is still releasing models, still scaling compute, still moving the ball forward. But there's a difference between "we should go faster" and "we should think about whether we're going in the right direction." The second position is closer to what Pachocki seems to be taking.
It's a small thing. One essay by one scientist. But small cracks in the unity of the acceleration narrative matter. They're how conversations change direction.
Top comments (0)