“Which AI researchers moved between frontier labs since 2025? For each move list the researcher, the lab they left, where they went and the month.”
A conservative evidence screen over public reporting yields 39 researcher moves between frontier AI labs from January 2025 through October 8, 2026, drawn from 19 independent news hosts. This is not an exhaustive census — unannounced or thinly reported moves are not observable — and a broader, lower-confidence candidate set of 99 possible moves sits in the appendix below.
Ribbon width = number of moves on that corridor. Orange = flows into Meta Superintelligence Labs; grey = everything else. Hover a ribbon for the researchers behind it; click a month button or a bar to filter. Bars show moves per month — June and July 2025 peak at 8 each.
| Researcher | Left | Joined | Month | Sources | Source |
|---|
23 of the 39 rows are supported by a single independent host (marked “1 host”), even though the full set spans 19 independent hosts.
This wider screen keeps lower-confidence rows: single-source reports, affiliation-ambiguity cases and possible spelling duplicates (e.g. “Lachman”, “John Lachman”, “Jonathan Lachman”). It is shown for transparency and is not mixed into the headline totals — inspect the linked evidence before relying on any row.
| Researcher | Left | Joined | Month | Sources | Source |
|---|
Method: 39 conservative researcher moves extracted from 3,922 web search-result records across two year-targeted searches (before cross-search overlap), emphasizing Reuters, Bloomberg, Financial Times, Wired, TechCrunch, Fortune, The Verge and official lab announcements; moves normalized to lab brands. Window: January 2025 – October 8, 2026. “Month” is the reported joining or public-announcement month when an exact start month was unavailable. The conservative screen excluded advisers, rumors, stale historical affiliations, nontechnical executives, policy-only moves, robotics/hardware-only roles, and organizations not clearly pursuing frontier general-purpose or foundation models. A 99-row broad candidate set is appended unmixed; source titles were cut from tables for space.