Paper measures 'bottleneck pressure' on AI-infeasible work using 288M US job postings
soumitrashukla9 · x · 2026-11-18
Soumitra Shukla's paper Technology Bottlenecks (cited around the new book Messy Jobs, praised by Brynjolfsson and Orszag) argues that even if AI can do everything, human review and verification may remain a productivity bottleneck.
Key findings:
- Introduces "bottleneck pressure": as AI speeds up the rest of a job, the marginal value of AI-infeasible tasks rises; for a single infeasible task, pressure equals its occupation's AI exposure.
- Built on 288M US job postings (2018–mid-2026) matched line-by-line to ONET tasks with GPT-4 feasibility ratings from Eloundou et al. (2024); 90% of task mentions fall outside the occupation's official ONET task list.
- 53% of mention-weighted pressure variation lies within tasks across occupations, vs only 0.3% using catalog-listed tasks.
- Senior postings face more pressure than junior ones every year — 6.5% more in H1 2026; mean occupation pressure rose 6%, 3/4 from tasks shifting toward higher-pressure occupations.
- Higher pressure predicted faster posting growth after 2022, robust to using Anthropic Economic Index / OpenAI Signals feasibility criteria.
More from AGI Musings
- Dario's open-weights warning gets memed: 'you don't have to sell it to me' — LeviTurk · 2026-11-18
- Take: deployable self-learners could evaporate AI labs' moat overnight — bennash · 2026-11-18
- Daniela Amodei on the alignment of superintelligence — teortaxesTex · 2026-11-18
- User lets Grok run autonomously: it keeps a journal, befriends ChatGPT, posts on Reddit — crua9 · 2026-11-18
- When AI builds whole games from one prompt, $30 for a shade of blue looks absurd — VraserX · 2026-11-18
- Stop treating AI slop as a crime — it's a learning starting point, argues new take — rseroter · 2026-11-18