Lambert expects rapid AI gains, not general superintelligence

Rapid AI progress is coming, Nathan Lambert argues, but the destination is not general superintelligence. In a post on Interconnects dated October 9, 2026, he writes that 'the models will be superhuman distributed GPU engineers in a few years,' and that 'I expect the effective cost of model intelligence to decline near-exponentially over the coming years.'

The split he draws is between infrastructure and research. Coding agents make engineering cheaper, which lets labs run more experiments, and he calls pretraining research being automated in 2-3 years a reasonable guess. But that compounding, in his words, 'will not make our models dramatically different in nature.' A sharper ceiling follows: 'we don't get economically valuable superhuman traits outside of math and coding.'

Lambert points to Meta's Muse as evidence that near-term value comes from delivery rather than from new capability frontiers. 'Meta's Muse agent is an early indicator of this,' he writes, 'and we should expect more Muse-like experiences for different audiences and target use-cases.'

He opens by noting his surprise at hearing from 'top researchers in industry that they think AI will be better than them at their job in a few years,' and the piece reads as a counterweight to that framing without dismissing it. Reinforcement-learning data quality, he adds, is one lever that is 'clearly fixable.' The post names no specific lab timelines beyond Meta's Muse.