On November 30th, OpenAI released a “research preview” called ChatGPT, the earlier thread (GPT-3’s API demos) wrapped in a chat interface with RLHF post-training, and the fortnight since has been the fastest product adoption event in recorded history: a million users in five days, my non-tech family members sending me screenshots, and every group chat this archive feeds from conducting the same experiment simultaneously (write my performance review; explain my toddler’s sleep regression as a Nabokov passage; debug this regex, it did, the old backtracking monster, correctly, with an explanation). The staff-file assessment, written carefully because this one will be re-read: the interface is the revolution more than the model, GPT-3.5’s capabilities existed in API form for months (the four-day tools); wrapping them in dialogue with memory and refusal training converted a developer curiosity into a consumer epiphany, the Wordle lesson (distribution is design) at planetary scale. The failure modes are unchanged and now globally distributed (confident confabulation, fluency-as-failure-mode, meets a billion users who read fluency as authority; the citation-inventing lawyer incidents are pre-registered herewith), the academic-integrity panic has arrived on schedule (take-home essays met their classifier moment overnight), and our own office’s censused usage went from zero to “most engineers, daily” in ten days, boilerplate, tests, unfamiliar-API scaffolding, the fastest tool adoption I’ve witnessed in a career of platform evangelism (the funnel, self-executing; my job just changed and the file knows it). The quiet thread is no longer quiet, no longer a thread: as of this fortnight it is the loom. The 2023 file is pre-named.
The World Cup’s knockouts delivered on the earlier plot: Argentina survived (Messi in full late-career sorcery, the fourth-ring-weight doctrine building toward its ultimate test), Japan fell on penalties after topping Spain and Germany’s group (deprecation round complete, the Germans exit in groups again), Brazil got Croatia’d in tonight’s quarter (the 2018 file’s extra-time vampires, undead still), and Morocco, Morocco! toppled Spain on penalties and stand one Portugal quarterfinal from becoming the first African semifinalist ever, behind a defense of pure collective will and the tournament’s best story. The quarters conclude this weekend; the bracket’s dream endgame (Messi-Mbappé with Morocco playing spoiler) is two matchdays from possible, and the archive can barely file the narrative density.
TIL: RLHF mechanics, reward models trained on human preference rankings, then policy-optimized: the “helpful, harmless” shaping that made the interface feel possible. The alignment technique is the product breakthrough, and this entry will be cited for that sentence or mocked for it; the archive accepts both futures, that’s what the archive is for.