Weekly, from the scored takes. The line is the mean score of that week's takes; the bars are how many takes there were and how they split by level. Dashed rules mark the days with the most posts at 75+ (DEFCON 2 or worse).
Every post below scored 100 or 75 on the 0–100 alarm scale, written by someone whose own bio placed them at a frontier lab. People who have since resigned or been fired are marked. Highest score first, then reach.
Written from the ledger below. Every quoted line links to the actual tweet.
21-day rolling mean alarm (0–100) per lab, drawn only where a lab has at least four scored takes in the window. Gaps are honest: not enough data, not calm.
Headcount is the latest public figure. Tracked = accounts on X whose own bio names the lab today. Tweets per person are what we captured in the window, so they measure our coverage as much as their habits. "Alarming" means scored 75 or 100.
Left: headcount over time from public sources, nothing interpolated. Right: share of each lab's tracked accounts whose bio states a current safety, alignment, interpretability, evals or red-team role.
Two sources that don't depend on who posts. Left: every open role on each lab's own job board, classified by title — what share of hiring right now is frontier safety (alignment, interpretability, dangerous-capability evals, model welfare, preparedness). Right: the published field census of full-time safety researchers per lab.
Every move announced in a first-person post by someone we track, May to now, with the link. Safety-team moves are marked. The bar chart is each lab's net quarterly growth (hires minus departures) from LinkedIn-derived data.
Distribution of every scored take across the 0–100 scale.
Live Polymarket odds, snapshotted each run so this becomes its own time series. The employees' alarm and the market's prices are two independent reads on one situation.
Ranked by how many of their posts scored 75 or 100. Follower counts are small on purpose — this is not the executive layer.
Every scored take. Sort by alarm, date or likes, and narrow to a date range. Score is a judgement, one per tweet; the note says why.
Every account on the roster: people whose own bio says they work at a frontier lab now. Follow the whole set with one click on the X list, or one by one below.