FILE 00118KIn 2021, seven OpenAI employees walked out to found Anthropic, the lab people leave for. On September 8, a pretraining researcher resigned from Anthropic in a thread that passed sixty million views in a day, the company's alignment lead replied that he is right and put extinction odds above ten percent in ten years, and a safety researcher backed the warning in a personal capacity. Five years of resignation letters have become a genre with settled conventions, escalating costs, and no measurable effect on the race they describe. Studying why the warnings fail tells you what kind of restraint actually works, and why 1,178 frontier-lab employees are now asking the government to build a brake they cannot build themselves.
FILE 00214KAt the Stratford Festival this weekend, Yann Martel argued that the West's oldest stories endure because they cannot be verified: Troy, the Gospels, and in his telling, Donald Trump, who lies daily without cost because he speaks at the level of myth. It is a serious case for the uncheckable, and it names by opposition what this blog has been documenting all year: the moment a claim meets the layer where it can be checked. Writing was invented for accountants. The live question is which claims we still allow to live outside the ledger, and what happens when one crosses over.
FILE 00319KOn June 4, two days into a seventy-day training competition, the crown landed on a checkpoint that was more than half bit-identical to a model that already existed. A public audit has now shown what that means: the season's flagship from-scratch claim in crypto AI descends, weight by weight, from Alibaba's Qwen. The scandal is real. So is the inversion inside it: only open weights could have been caught this way, and the honor system for training claims ended the moment anyone could check.
FILE 00419KA vehicle-routing method beat the state of the art this year and was published for anyone to read. Who owns it? TIG's licensing document is the fullest answer crypto has produced to the value-capture question: copyright conceded dead, patents on technical effects, a data-disclosure license that herds commercial users toward payment, and community enforcement in place of courts. The mechanism is real. The demand is unproven. Here is what would have to be true.
FILE 00527KI did not lose faith in open AI, open science, or decentralized infrastructure. I lost faith in the assumption that every important coordination problem deserves an investable token. That was the economic reckoning. Then five mornings at Stratford added the political one: every registered claim has an owner, the owner takes sides in an emergency, and the strongest monetary-economics argument of the decade — Alden's — quietly assumes the emergency never comes. This is the essay the authoritarianism series kept promising: what a bearer asset is actually for.
FILE 00629KThe fifth and final Stratford panel was about no dictatorship in particular, which means it was about us. Three scholars described an authoritarianism that arrives by election and governs by paperwork, then did something no panel all week had tried: they ran the stages of fascism backwards and built a checklist for dismantling one. Along the way they stress-tested this series' own arguments, defended the most unfashionable people in any democracy, and never got to the two questions I wrote on an index card. This essay answers the card.
