What Will Be Left for Us to Work on? – Key Takeaways from the ICML 2026 Keynote
What Will Be Left for Us to Work on? – Key Takeaways from the ICML 2026 Keynote
TL;DR
The speaker’s three core claims are:
- AI as a normal technology – AI will behave like past transformative tools (electricity, the internet) unless a sudden breakthrough such as recursive self‑improvement (RSI) occurs.
- No imminent “job‑killing” milestone – Even if RSI is achieved, it will not instantly replace human workers; economic impact unfolds over decades.
- Future work will be about evaluation, judgment, and augmentation – As AI takes over verifiable tasks, human value will concentrate on steering, assessing reliability, and applying creativity.
1. AI as a Normal Technology – A Proven Historical Pattern
"AI as Normal Technology is the intellectual framework for my talk today. When we say AI is normal, we don’t mean that it’s just like a hammer or a toothbrush… It’s a transformative technology on the scale of the industrial revolution." – Arvind Narayanan
The framework mirrors the diffusion‑of‑innovations model used for electricity:
- Invention – discovery of core principles (e.g., electromagnetism, LLM architectures).
- Innovation – creation of usable products (coding agents, AI‑assisted legal research).
- Diffusion – gradual adoption across industries.
- Adaptation – slow, decades‑long re‑organization of work, institutions, and regulations.
The speaker stresses that adaptation is the slowest phase and has barely begun even in early‑adopter fields like software engineering.
Comment Insight
"The same pattern happened with ATMs and radiology – automation made the service cheaper, which increased employment rather than eliminated it." – HN comment (Metricon)
This historical evidence supports the claim that AI will likely expand demand for skilled labor rather than create a permanent underclass.
2. Recursive Self‑Improvement (RSI) Is Not a Shortcut to Mass Unemployment
The talk distinguishes three possible meanings of RSI:
- Hyper‑parameter search – automated tuning, already common (AutoML).
- Full‑scale research automation – AI replaces the creative work of thousands of researchers.
- Hybrid – AI improves efficiency of verifiable tasks (speed, cost) while human creativity remains the bottleneck.
The speaker argues that creativity, judgment, and representation quality are still far from human level. Evidence:
- Capability (accuracy) has risen dramatically in the last 24 months.
- Reliability metrics (consistency, robustness, calibration, safety) have improved by only 5‑10 percentage points over the same period.
- Current benchmarks conflate raw accuracy with reliability, obscuring the true deployment gap.
Comment Insight
"He’s basically saying that even though AI capability is high and rapidly increasing, it is not reliable, creative or tasteful enough to replace humans…" – HN comment (ilaksh)
The community’s concern aligns with the speaker’s emphasis on reliability as the primary barrier to automation.
3. The Future of Work: From Building to Evaluating
3.1 The “Decide‑Execute‑Deliver” Sandwich
- Decide layer – problem definition, requirements, planning – not compressible by AI.
- Execute layer – coding, debugging – compressible, but historically only ~⅓ of the work.
- Deliver layer – integration, testing, accountability – also not compressible.
When the execute layer shrinks, the decide and deliver layers expand, shifting labor toward higher‑level coordination and judgment.
3.2 The Evaluation Discipline
- AI‑agent evaluation has become a distinct, automation‑resistant field.
- Evaluation is analogous to steering a ship: the engine (AI) provides power, but humans decide direction.
- The speaker predicts evaluation could occupy up to half of AI conference content within a few years.
Comment Insight
"Work is shifting from building/doing to evaluating, judging, and steering — that’s where human value will concentrate." – HN comment (chopete3)
This mirrors the broader trend across professions: lawyers, translators, and radiologists see new work generated by AI‑enabled demand (e.g., more lawsuits, more translation requests), not fewer jobs.
4. Practical Recommendations for Individuals
- Invest in complementary, non‑verifiable skills – judgment, domain expertise, and the ability to design AI‑augmented workflows.
- Avoid the “black‑box trap” – understand model outputs, calibrate confidence, and retain control.
- Follow the “dependence spiral” heuristic – master a task before delegating it to AI; otherwise skill erosion occurs.
- Reinvest productivity gains – allocate time saved by AI to learning new domains, not just short‑term output.
5. Policy and Governance Implications
- Regulation must evolve to address reliability, safety, and tacit knowledge integration rather than merely restricting model releases.
- Evaluation frameworks can act as a form of alignment for the research community, ensuring AI development follows socially desired trajectories.
- Long‑term strategic planning should treat AI as a tool that amplifies human intelligence (co‑superintelligence) rather than a replacement.
6. Community Reactions on Hacker News
| Theme | Representative Comment |
|---|---|
| Skepticism about total automation | "What will be left for us to work on? Nothing. The wealthy own all of the infrastructure…" – Suzuran |
| Optimism about new roles | "There will always be new, hard problems to work on. AI will not eliminate that." – jppope |
| Concern over AI fatigue | "I have AI‑fatigue… I’m wondering if it might hit more fields and cause humans to reject AI work." – doubtfuluser |
| Analogy to medical professions | "Software developers will become like doctors, with nurses and medics supporting them." – Metricon |
| Emphasis on evaluation | "Work is shifting from building to evaluating, judging, and steering…" – chopete3 |
| Questioning the timeline | "Reliability isn’t at zero and is increasing rapidly; the claim of decades‑long lag feels speculative." – ilaksh |
These comments illustrate the spectrum of reactions: from alarmist pessimism to measured optimism and practical concerns about how AI will reshape professional identities.
7. Concluding Vision – Co‑Superintelligence
The speaker ends with a metaphor:
"Computers have often been called bicycles for the mind. I think AI can be a crane for the mind, amplifying our potential to previously unimaginable heights."
The co‑superintelligence vision posits a future where humans operate, understand, and direct increasingly capable AI systems, preserving agency while leveraging AI’s productivity gains.
Key Takeaways
- AI will remain a collaborative amplifier for decades, not an instant job‑killer.
- Reliability, not raw capability, is the current bottleneck; progress on safety and evaluation will dictate adoption speed.
- Human work will shift toward evaluation, judgment, and creative orchestration—the “steering” of AI‑powered systems.
- Policy, education, and institutional adaptation are essential to ensure the transition benefits society rather than concentrating power.
- Individuals should focus on mastering high‑level, non‑automatable skills and maintain control over AI tools to avoid skill erosion.
The full slide deck and annotated transcript are available at the speaker’s Princeton page.