Princeton's Narayanan: No Lab Milestone Will Suddenly Automate Your Job

In a keynote at ICML 2026 in Seoul, Princeton researcher Arvind Narayanan pushed back on the anxiety gripping the AI community over its own obsolescence. His core argument rests on the “AI as Normal Technology” framework he developed with Sayash Kapoor: AI is genuinely transformative on the scale of the industrial revolution, but its economic effects will unfold through the same slow, four-stage process as past general-purpose technologies — capabilities, then products that package them (coding agents, not raw LLMs), then early adoption, then structural adaptation. That final adaptation phase, he argues, takes decades and has barely begun even in software engineering, the field adopting AI fastest.

Narayanan takes recursive self-improvement seriously as the one scenario that could break this model, but he doesn’t lose sleep over it, and he rejects the idea that any single benchmark result or lab achievement will abruptly render human work obsolete. He frames the moment as a fork between two bets: treat AI as a replacement technology and rush to build wealth before your skills become worthless, or treat it as an amplifier and invest now in complementary capabilities — agency, taste, judgment. He argues the second path is the safer wager, since choosing wrongly on the first means squandering what may be the best window in history to build skills that compound with AI.

He also warns that how technologists respond carries political weight: passively conceding work to AI rather than setting boundaries risks a sharper public backlash. Looking ahead, he speculates that reliable coding agents could shift software away from mass-market products toward “extreme personalization” — software tailored per person or team — and even question whether standalone software companies remain necessary as development moves in-house. He closes on a vision of human/AI “co-superintelligence” rather than human replacement.