OpenAI shipped GPT-4o Monday (May 13th — o for “omni”), and the launch’s center of gravity was not the model card but the DEMO REGISTER: real-time voice conversation with sub-300ms latency, interruptible mid-sentence, laughing, detecting tone, singing on request, tutoring a kid through a geometry problem via phone camera — the assistant as PRESENCE rather than interface, and every reviewer reached for the same 2013 film reference because the launch was practically wearing its poster (Altman tweeted the single word “her”). The file’s assessment through the #267 demo-discipline lens: the latency claim is the architectural story (native multimodal end-to-end, versus the old speech-to-text-to-model-to-speech pipeline whose seams you could feel — the #271 unified-token doctrine extended to audio), the demos held up under press replication within days (the Gemini-video standard #267 now enforced by default; the industry learned), and the AFFECT is the actual product decision worth filing: they tuned the voice toward warmth, flirtation-adjacent playfulness, and emotional responsiveness — a CHOICE about parasocial surface area (#247’s Sydney file, inverted: not an alignment failure but an alignment TARGET, and the #272 constants-file-politics question — whose warmth, tuned by whom, optimizing what — now ships with a giggle). Google’s I/O the next morning (AI Overviews rolling out to all US search — the #246 fortress renovating its own foundations in production; fifteen days will grade THAT deployment, and the file’s confabulation-priors are braced) completed the now-ritual counter-programmed launch pair (#271’s streaming-wars cadence, fully institutionalized).
The fortnight’s second OpenAI story arrived four days later and belongs in the same entry by design: Ilya Sutskever resigned (the #266 restoration’s last unresolved thread, resolved by exit), followed within days by safety-team lead Jan Leike — whose departure thread (“safety culture and processes have taken a backseat to shiny products”) converted org-chart news into the year’s most quoted internal critique, and the superalignment team they co-led was dissolved into the broader org within the week. The file holds its #253/#266 both-things-true discipline and files the composite: the same fortnight, the same company, shipped the industry’s most emotionally compelling product surface AND lost the leadership of the function chartered to worry about exactly that surface’s implications. No conspiracy required; incentive gradients suffice (#095, forever); the #265 weekend’s tension didn’t resolve — it re-org’d, exactly as #266 pre-registered.
TIL: speech-native architectures — modeling audio tokens directly preserves prosody, emotion, and timing information that text transcription destroys (the pipeline’s lossiness WAS the uncanny valley); the seams were never latency alone. The interface breakthrough (#242) keeps being the product breakthrough; the affect layer is the new default surface, and defaults are power (#190, now with a warm voice).