AI researcher says one hard task can block full automation of entire jobs
herbiebradley · x · 2026-07-24
In a repost of an interview clip, AI researcher Herbie Bradley explains the O-ring problem: one hard-to-automate task can block automation of an entire job.
He argues that reinforcement learning is strong at highly verifiable, grindable tasks such as coding and AI R&D, but much weaker on qualitative work. Since most jobs combine verifiable and non-verifiable tasks, automation tends to get stuck on the few harder-to-automate pieces, often involving human relationship building.
He frames this as an O-ring dynamic from economics: like the Challenger shuttle disaster, a single failure point can determine the outcome of the whole system, so full job automation may remain bottlenecked by a small number of difficult tasks rather than the bulk of the workflow.
More from AGI Musings
- Dev Rant: If AI Coding Agents Mess Up, Why Are We Paying for the Tokens? — larumis · 2026-07-24
- A 2016 UDT result was already shown in 1997, author says — jessi_cata · 2026-07-24
- Frontier AI labs are shifting their messaging on jobs and AI impact — menhguin · 2026-07-24
- Scott Galloway says cheaper Chinese open-weight models could break US AI valuations — r0ck3t23 · 2026-07-24
- From Builders to Debuggers: The Shift in AI-Era Software Development — zakelfassi · 2026-07-24
- AI labs sell recursive self-improvement because 8% progress sounds too small — joshalbrecht · 2026-07-24