Four days of old vibe coding

The market check first: tools that clip videos exist, but they bill by the minute, one plan selling 100 hours a year when this channel produces that in ten days ▸ 1:55, pricing Julia recognizes instantly as the limited-minutes phone plans of their childhood ▸ 2:12. So Snipy continues, and the log is the entry.

Day one: the stereo idea dies again, the mics are mono. But the search term it surfaces, diarization, turns out to be a mature research field, and Pyannote segments the audio into speaker timelines the first night ▸ 4:16. The MVP identification step is pure detective logic: three detected speakers collapse to two plus filler, and “yo siempre hablo más, entonces yo soy speaker 02” ▸ 5:59, left and right assigned by hand, the demo cutting between them on camera ▸ 6:23. Then ambition: an active-speaker-detection paper, AVA-AVD, boxes around faces, in theory re-identifying a person who leaves the frame ▸ 7:09. The catch consumed two days: the GitHub repo is research exhaust, published as-is when the paper shipped, “tome, le botó eso ahí”, tuned for the author’s GPUs and 200-video datasets, not for a stranger with one video ▸ 8:59.

cuando el agente da vueltas, se vuelve al viejo oficio: un modelo piensa, otro teclea →

The tooling story inside the sprint: Gemini 2.5 Pro looped uselessly once the project grew too many files ▸ 10:16, so he regressed to “old vibe coding”, one model as architect, ChatGPT, whose instructions he pipes into Cline for execution ▸ 14:24, burning the free tier across three accounts ▸ 15:26. The actual unblock was returning to “la vieja confiable”, Claude, Sonnet 4.5, which he hadn’t touched in ages and which shocked him with how much it had improved ▸ 16:05, worth four exhausted accounts and a promise: the day there’s money, that subscription gets paid ▸ 18:37. State at deadline minus one: every face currently labels speaker zero ▸ 19:36, a parallel agent picks the clip-worthy transcript sections, clickbait welcome, under three minutes ▸ 19:55, and the mature call is already made: he won’t finish in time, the assessment grades process anyway, and he’s building the app regardless, “la quiero hacer para mí” ▸ 21:26…

Postscript, two days later: the bounding boxes were misfiring and he was ready to swap models; the newly paid-for Claude suggested tuning the detection threshold instead, he called it a stupidity, ran it, and the tracker started following person A and person B “casi que perfectamente” ▸ 54:01. Remaining bugs are honest ones, identities drift to person D and E when someone leaves frame, screen-shares explode the roster, and a 40-minute video crashed the computer at minute 25, but the itch is back: “me dan ganas de volver a trabajar con machine learning” ▸ 55:54.