DEVOURED
What will be left for us to work on?
Arvind Narayanan argues that human labor is shifting from direct execution to 'evaluating' AI systems, which he terms 'co-superintelligence.'
What: In his ICML 2026 keynote, Princeton researcher Arvind Narayanan posited that recursive self-improvement will not cause sudden mass unemployment. He advocates for prioritizing 'eval' over 'build' roles and argues that AI remains a tool for collaboration rather than a drop-in replacement for cognitive workers.
Why it matters: This reframes the AI anxiety narrative by suggesting that economic transformation will take decades rather than months, and that human judgment in evaluating AI output is an essential, non-automatable skill.
Deep dive
- AI acts as 'normal technology' on the scale of electricity, requiring structural societal adaptation.
- Current benchmarks over-index on 'accuracy' while failing to measure 'reliability' (consistency, robustness, safety).
- The 'decide-execute-deliver' framework shows that coding is only one-third of the software development lifecycle.
- Knowledge workers should focus on 'operating' machines rather than competing to perform the cognitive heavy lifting themselves.
- AI creativity currently fails at verifiable scientific reasoning and representation building.
Decoder
- Recursive self-improvement (RSI): The theoretical ability of an AI system to iteratively improve its own source code and architecture, potentially leading to an intelligence explosion.
- Jevons' paradox: A situation where efficiency gains in resource usage lead to higher total consumption, rather than lower, due to increased demand.
- Lump-of-labor fallacy: The economic misconception that there is a fixed amount of work to be done, meaning automation must decrease total employment.
Original article
Full article content is not available for inline reading.
Explore all 41 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi