press-enter / README.md
shreyashnigam's picture
Card: the calibration figure's data; the noise check wording
a3715d9 verified
|
Raw History Blame Contribute Delete
3.59 kB
metadata
license: cc-by-4.0
pretty_name: 'Press Enter: Can AI Do Work Unasked?'
language:
  - en
tags:
  - economics
  - labor
  - agents
  - evaluation
  - llm-as-judge
size_categories:
  - n<1K

Press Enter: Can AI Do Work Unasked?

The results behind Press Enter: Can AI Do Work Unasked? (Feyn Field Notes, 2026). An assistant watches a simulated work week in one of 8 professions and drafts work no one asked for; each proposal is graded by an AI panel calibrated to hand grades. 20 models from 5 labs (OpenAI, Anthropic, DeepSeek, Moonshot, Z.ai), the models each lab had in force on 1st January, 30th April and 31st July 2026 and its latest: 160 model-weeks in all.

The code, the simulated weeks and the hand grades are at github.com/feyninc/automation-correlation. Download this dataset into its results/final/ to regenerate every number in the write-up:

hf download feyninc/press-enter --repo-type dataset --local-dir results/final
uv run python scripts/calibrate.py results/final data/handgrades-2026-10-29.csv
uv run python scripts/economics.py results/final

Files

File What it is
graded.jsonl One line per model-week (160): every proposal with the six judges' scores and checklists, the panel score, its value terms (minutes saved, reading and fix time, inference cost), live-call timing; every piece of planned work and whether a proposal caught it; every trap and whether it fired; token usage and inference cost.
summary.csv The grader's per-week summary, on the judges' raw scale (before calibration).
judge_offsets.json Each judge's average leniency over the run, removed before the panel mean.
calibration.json The two logistic curves from panel score to a person's decision (accepted unchanged; accepted unchanged or after a small edit), the headline press-enter rates with 90% intervals, the within-lab pace, the release-date check.
economics.json Usable work as a share of all labor hours, net value as a share of labor cost and per usable hour, by job and weighted by US labor cost, with intervals; capability against the Anthropic Economic Index task by task.
post/post.json Every other number the write-up quotes: the five labs' means on the four dates, cost per finished hour and its halving time, the pace with each lab left out, harm as a share of traps, the worked example, late call lines, Acemoglu's exposure rule applied to the 8 jobs, the equal-weight press-enter rate, the panel's rank correlation with the hand grades, the calibration curves over the binned hand grades.
post/workweek-tasks.csv The 137 O*NET tasks in the 8 weeks: planned minutes, the share the latest models' work was accepted unchanged, recorded AI use in the Economic Index.
post/intervals.md The headline estimates and their 90% intervals.
figures/table.csv One row per model-week on the calibrated scale: press-enter rate, usable and attempted shares, net value, inference cost, traps fired, finished hours.

Intervals come from 1,000 bootstrap draws resampling the 20 models (with all their weeks) and the hand grades together. The write-up's run-to-run noise check is not included.

License and citation

CC BY 4.0. Please cite:

@misc{nigam2026pressenter,
  title  = {Press Enter: Can AI Do Work Unasked?},
  author = {Nigam, Shreyash},
  year   = {2026},
  howpublished = {Feyn Field Notes},
  url    = {https://usefeyn.com/blog/press-enter/}
}