Nathan Lambert expects fast AI progress without a leap to superintelligence

Nathan Lambert thinks the next few years of AI will be fast but not mysterious. In a new essay he argues the gains will come mostly from automating engineering work, not from models becoming a fundamentally different kind of thing. Coding agents make experiments cheap to run, which pushes the bottleneck back toward research: when execution is easy, good ideas matter more than good implementation. He prefers to call this "parallelized, AI-assisted language modeling" rather than recursive self-improvement, since the latter smuggles in takeoff assumptions he doesn't share.

The concrete version of his claim is that models will become superhuman "distributed GPU engineers." Training speed and inference cost are highly verifiable targets: tokens per second per GPU, cost per answer, FLOPs per token. Because those metrics are easy to measure, agents can optimize them end to end, and Lambert expects the effective cost of intelligence to keep falling at a near-exponential rate, possibly faster than it has recently. Co-designing chips and models over longer horizons could add further orders of magnitude. He puts automating pretraining research, meaning the choices of architecture and data, at a reasonable two to three years away.

What he is careful not to claim is a jump to general superintelligence. Progress in science, he notes, might surface useful cross-field connections in biology and chemistry, but whether that cures most cancers or just speeds up existing trends is genuinely unclear to him. Demand for agents will rise sharply, and the hard part becomes orienting and delivering them rather than raw capability. You can read the full argument on Interconnects.

Why it matters

If you plan around AI, Lambert's framing says to bet on much cheaper inference and heavy automation of engineering work, while staying skeptical of promises that a single model will suddenly become generally superhuman.