Year twelve opens with the copyright case this archive pre-registered fifteen months ago (“the litigation defining this decade files within months,” late by a year, correct in kind): The New York Times v. OpenAI and Microsoft, filed December 27th, and it’s the strong version of the claim, not just “you trained on our archive” but exhibit after exhibit of GPT-4 reproducing Times articles near-verbatim under adversarial prompting (the memorization receipts the fair-use debate has been waiting for), plus the market-substitution argument (Browse-with-Bing summarizing paywalled recipes) that turns an abstract doctrine into a revenue chart. The file’s read: this is the case both sides arguably want, the labs need the training-data question settled at precedent altitude rather than by a thousand district-court paper cuts, the publishers need leverage for the licensing market that is obviously the endgame (Axel Springer and AP already signed; the Times sued after negotiations stalled, which tells you the suit is the negotiation, contracts-as-adversarial-runbooks, media edition). Pre-registration for the file: no verdict ever lands, settlement plus licensing regime within two years, with the memorization exhibits driving the price. Either way, 2024’s AI story adds a fourth branch of government to the governance stack: the judiciary has entered the training loop.
The year-ahead board, per the earlier sprint list: an election cycle running on the degraded information immune system (the file’s braced posture holds), Vision Pro in February (the window opens), the agent watch-item compounding quietly, and, the macro ledger, rate-cut expectations doing to 2024 planning what hike expectations did to 2022’s (the dot-plot dependency, now bidirectional). CES next week will produce one genuinely odd artifact worth filing (early leaks suggest AI-in-everything again, plus a small orange box called the Rabbit R1 that the keynote circuit is inexplicably excited about, the file pre-registers Juicero-pattern skepticism pending the hand-squeeze test).
Sabbatical dispatch, year-opening (the break, now real): the calendar is empty for the first time since 2013, and the M.Tech coursework has expanded to fill it, neural-network assignments due the same month the frontier makes them feel like archaeology, which is itself an education in curve-slope. And a report from the old badge, via the ex-colleague group chat that keeps me honest: the review-first curriculum graduated its first full cohort this month without me, their code-review comments sharper than any prior cohort’s at the same tenure. The ladder holds in my absence, which was always the actual test: the curriculum turned out to be the deliverable, not the teacher. The Proverbs file opens year twelve at 301 entries and one thesis: judgment compounds, tools amplify, minutes matter, the closing wager, now the operating system.
TIL: memorization-vs-generalization measurement, extraction-attack methodology, dedup’s role in training pipelines, and why verbatim regurgitation is an engineering artifact (data repetition in the corpus) as much as a legal one. The exhibits are a training-pipeline code review conducted by opposing counsel (the discovery doctrine: the industry’s only true documentation, now including the datasets).