Press Enter: the run's results (160 model-weeks), calibration and the write-up's numbers
958ce81 verified |
Download post/intervals.md from feyninc/press-enter: direct link, hf CLI and curl.
- Browser
- Download file 1.23 kB
-
https://huggingface.co/datasets/feyninc/press-enter/resolve/main/post/intervals.md
- Command line
-
hf download hf://datasets/feyninc/press-enter/post/intervals.md
-
curl -L -o intervals.md https://huggingface.co/datasets/feyninc/press-enter/resolve/main/post/intervals.md
1.23 kB
Headline estimates and 90% intervals
Intervals: 996 bootstrap draws resampling the 20 models (with all their weeks) and the hand grades together.
| Measure | Models of 1st January | Latest models |
|---|---|---|
| Press-enter rate (accepted unchanged, share of planned hours) | 7.1% [3.2%, 12.4%] | 19.6% [13.0%, 26.7%] |
| Usable (unchanged or after a small edit), share of planned hours | 20.9% [10.9%, 32.9%] | 49.5% [38.9%, 57.6%] |
| Latest over January press-enter rate | 2.76x [1.58, 5.84] | |
| Pace within a lab, press-enter rate | 16.1 points a year [9.9, 23.0] | |
| Usable work, share of all labor hours (labor-cost weighted) | 9.1% [4.7%, 13.9%] | 20.4% [16.4%, 23.4%] |
| Finished hours, share of all labor hours (labor-cost weighted) | 3.1% [1.5%, 5.2%] | 7.7% [5.1%, 10.3%] |
| Net value, share of weekly labor cost (labor-cost weighted) | -2.6% [-12.0%, 4.8%] | 12.4% [9.0%, 15.0%] |
| Net saving per usable hour | 60.8% [54.6%, 65.3%] | |
| Multiple of the 1.24% Acemoglu's economy-wide inputs imply | 10.0x | |
| Multiple of his inputs for the same 8 jobs (2.4%) | 5.1x [3.7, 6.2] | |
| Hand-graded calibration | n = 92 distinct proposals | release-date term 0.727 log-odds a year [-1.116, 3.422] |