OpenAI’s #266 re-org completed its slow-motion arc this week: Mira Murati resigned Wednesday (the interim-CEO of the #265 weekend, the product org’s center of gravity), followed within HOURS by the chief research officer and a research VP — the same week reporting confirmed the company’s restructuring toward removing nonprofit control entirely (the capped-profit wrapper #245, whose stress test #265 revealed the cap table’s actual power, now being formalized into the org chart it always was — with equity stakes for the CEO under discussion, per reporting the company disputes in emphasis). The file’s ledger of departures since the #265 weekend now reads: Sutskever, Leike, Karpathy, Schulman, Brockman-on-leave, Murati, McGrew, Zoph — functionally the entire founding research and safety leadership, out within ten months of the board’s capitulation, and the #266 extraction (“the tension didn’t resolve — it re-org’d”) upgrades to its terminal form: THE STRUCTURE RESOLVED BY EXIT. Whatever one’s read on any individual departure (startup attrition is real; so is gradient — #277), the aggregate is a governance postmortem written in resignation letters, and the archive files it next to #164 with the observation it has earned across twelve years: incentive structures don’t fail loudly; they fail by SELECTION — the people whose concerns priced above their equity simply leave, and the org that remains is, definitionally, the org that didn’t share them (#095’s Goodhart, applied to workforce composition; the metric survived, the mission migrated).
The same weekend, California’s SB 1047 — the frontier-model safety bill (liability for catastrophic harms, shutdown capabilities, compute thresholds) that split the industry down novel lines (Anthropic cautiously-for after amendments; OpenAI, Meta, and most of VC-dom against; Musk, chaotically, FOR) — was VETOED by Newsom, whose message argued the compute-threshold approach regulates the wrong variable (size over deployment risk — the #264 10^26-half-life problem, now a veto rationale) while promising empirically-grounded regulation instead. The file’s read: both the bill’s champions and its critics have half the truth (thresholds ARE crude; “wait for evidence” IS how every #146 invoice gets financed), the vetoed text will be back in amended form within a legislative cycle, and the real signal is jurisdictional — the #253 regulatory-geography map now includes SACRAMENTO as a frontier-AI capital, because the labs all live there and Brussels-effect logic (#131) works at state scale too (see: CCPA).
The fortnight’s hardware footnote with a decade’s fuse: Meta demoed ORION — genuine augmented-reality glasses (waveguides, neural wristband input, actual field of view) shown as prototype, not product — and the #215 $10B-a-year burn finally displayed the thing it was buying: the file, on record doubting the metaverse framing since the rename, notes without irony that glasses-that-work were ALWAYS the endgame worth the burn (#254’s Vision Pro pre-registration meets its future competitor; the form factor war of the 2030s got its first honest preview), and Hurricane Helene’s devastation across the Southeast (filed with #259’s care: the climate-infrastructure ledger grows again — mountain towns flooded beyond design basis, the #233 envelope broken in a new geometry) closes the fortnight with the archive’s recurring prayer: fund the plumbing, believe the projections, rehearse the response.
TIL: nonprofit-to-PBC conversion mechanics — asset-valuation requirements when a charity’s assets (here: control of the decade’s most valuable startup) transfer to for-profit form; attorneys general get a say, and the price of the mission is, for once, a literal appraisal question. The #245 governance experiment ends as an M&A transaction; the file has never typed a sadder TIL.