Mercor compares 29 AI benchmarks, fixed sets of tasks for comparing models, with the US Labor Department’s catalog of work. Its estimate: 90% of American working hours fall on tasks with little or no coverage in that selection. That is a gap in measurement, not a finding that AI can or cannot do those jobs.
- A task counts as covered when at least five test prompts map to it. Only 208 of 923 occupations have even one task meeting that threshold.
- Coverage favors occupations with larger total wage bills, even after accounting for how exposed their tasks are to AI automation. Lawyers get more attention than legal secretaries or court clerks.
- The selection leaves out tests models already perform very well on. Mercor acknowledges that its incomplete list can understate coverage.
- Models help produce the estimate itself: an LLM maps prompts to work tasks, and LLM-derived estimates supply the time spent on each task.
The analysis also asks whether tests of a model working alone can tell us how much it helps a person. Measuring assistance remains an open problem. Before using a model’s score to judge its value at work, ask which parts of the job the test actually represents.
The 291-comment thread on Hacker News adds objections to the coverage threshold and accounts of what happens after coding gets faster.
What the thread adds
- OtherShrezzing — “It’s self-evident that the spectrum of work carried out by a social worker undertaking that work cannot be well represented by thirteen prompts.”
- bunderbunder — “automating what we could automate made things worse overall.”
- legitster — “My time is now filled with things like presentations, Jira, QA, etc.” kfarr, replying: “QA / code review is like 80% of my time now! Like coding now is QA. That’s it.”
- cyanydeez — “I love the fact that I can finish programming something, then tell the AI to do the build/debugging and move on.”
Handles are pseudonymous; HN publishes no per-comment scores and uses its own ranking, so this is a slice of the thread, not a consensus.