Nathan Lambert: Expect rapid AI progress, but not toward general superintelligence

natolambert · x · 2026-10-10

Nathan Lambert lays out his view on AI progress: he believes AI will become far more efficient — a massive lever on diffusion — but does not believe models will make breakthroughs that eliminate all known weaknesses of LLMs.\n\n- Hearing top industry researchers say AI will beat them at their jobs in a few years, he initially doubted it. The resolution: those researchers see massive acceleration in the models' infra and engineering capabilities. He fully agrees models will be superhuman distributed GPU engineers within years, making experimentation on model formulation far easier — but this won't change the models' fundamental nature.\n- Ilya was early in declaring AI's "era of research" — with coding agents, research is indeed easier now, signaling a shift where good ideas matter more than good execution in software.\n- The research-vs-engineering balance has swung before: pre-deep-learning AI was research-heavy; today top researchers are judged by implementing and scaling ideas on complex infrastructure. A coming acceleration shifts the bottleneck away from engineering again — some call it RSI, but a grounded framing is "parallelized, AI-assisted language modeling." You don't need takeoff scenarios to see how infra improvements will change the nature of AI work.

Original post →

Nathan Lambert: Expect rapid AI progress, but not toward general superintelligence

More from AGI Musings