I run a job board covering one discipline: evals, post-training, red teaming, alignment and human data. It collects daily from 94 companies' own career endpoints.
Because I keep listings after they disappear, I can measure something I have not seen published anywhere, which is how long a role in this field actually stays open.
Across 500 roles that have opened and closed on the board: Median time open: 46 days. A quarter close within 19 days; a quarter are still up after 110.
By discipline, the spread is wider than the median suggests:
- Post-training and RL: 19 days
- Alignment and interpretability: 41 days
- Evals and benchmarks: 47 days
- Red team and safeguards: 52 days
- Human data and annotation: 74 days
I think the post-training number is the interesting one. A median of 19 days means most of these roles are filled from a pipeline the employer already had, and that a posting is closer to a formality than an invitation.
If you are watching for one, watching weekly is too slow. What this does not measure: I record a role as closed when it leaves the employer's own feed. That usually means filled, sometimes withdrawn, occasionally a careers-site migration.
It is a decent proxy for how long you have to act, and it is not a hiring statistic. Roles with a close date before their posting date, or open more than two years, are excluded as bad source data.
A second finding, on 623 currently open roles: only 34% publish a salary band at all. Among those that do, the median band is $210k to $325k. Everything is free and CC BY 4.0, as JSON and CSV, rebuilt every morning, with the method written out: https://aievalsjobs.com/data/
>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
Disclosure: this is my project. The data is free and stays free; I make money, in theory, from employers paying to feature a listing. I am posting it because the time-to-close numbers seemed worth having in the open, and because I would like to know whether anyone finds an error in them.
