What will be left for us to work on?

Newsletter · ICML 2026 Seoul keynote · Read

• Reframes “AI as Normal Technology” against sudden-displacement fears: AI follows historical technology-adoption patterns (like electricity) — four gradual phases (capability improvements, product innovation, early adoption, structural transformation) unfolding over decades, not months.
• Names the actual bottleneck as a reliability gap, not a capability gap: accuracy has improved dramatically but consistency, robustness, calibration, and operational safety lag, keeping automation confined to lower-stakes, human-collaborative applications.
• Separates four distinct, non-entailing progressions — recursive self-improvement, humanlike AGI, economically transformative AI, and superintelligence — arguing achieving one doesn’t guarantee the others, with creativity flagged as a specific barrier requiring better representations.
• Argues human effort shifts from building to evaluation as AI absorbs verifiable technical tasks — judgment work (deciding what to build, evaluating systems, steering direction) becomes the human role, analogous to ships needing navigators rather than rowers.
• Proposes “co-superintelligence” as the target: humans and AI augmenting each other rather than AI replacing humans, requiring deliberate resistance to black-box automation and maintained domain mastery before AI-assisted leverage.