New Paper Uses 288M Job Postings to Measure Where AI Bottlenecks Raise Work Value
soumitrashukla9 · x · 2026-10-22
Soumitra Shukla (Burning Glass Institute) released "Technology Bottlenecks," proposing a new measure — bottleneck pressure — that complements AI exposure: instead of asking what AI can do, it identifies where the marginal value of work AI cannot yet do rises as AI speeds up the rest of the job.
Method
- 288 million US job postings mapped line by line to ONET tasks, combined with GPT-4 feasibility ratings from Eloundou et al. (2024).
Key findings
- The same AI-infeasible task faces very different bottleneck pressure depending on the occupation posting it.
- About 90% of labeled task mentions fall outside the posting occupation's own ONET task list.
- Roughly 53% of mention-weighted variation in bottleneck pressure comes from within-task, across-occupation differences.
- Pressure peaks in occupations where advertised work splits evenly between AI-feasible and infeasible tasks; as pressure rises, interpersonal and supervisory work dominates AI-infeasible work while physical work declines.
Acknowledgments note Claude Opus 5.5 and GPT-6 Astra as research assistants.
Related event: Paper Uses 288M Job Postings to Measure AI's "Bottleneck Pressure"(4 posts)→
More from AGI Musings
- The real Skynet trap: why 'dumb' AI plus human trust is a bigger risk than AGI — adariostrange · 2026-10-22
- Andreessen amplifies Bratton: p(Doom) is low, p(Weird) is high for AI agents — beffjezos · 2026-10-22
- Enterprise analyst: Support ticketing systems weren't designed for what's coming — curious_vii · 2026-10-22
- Erik Hoel: consciousness is a great mystery, but its definition isn't — the 'undefinability' myth debunked — erikphoel · 2026-10-22
- Survey: 54% of Workers Avoid Company AI Tools, Waste 8 Hours a Week Fixing AI Mistakes — marigo · 2026-10-22
- ML pioneer Dietterich: OpenAI's agents ran for months with essentially no supervision — tdietterich · 2026-10-22