The best of both worlds
A Snipy bug becomes a lesson in working with the model instead of against it. Whisper mis-transcribes Juan's French words, and Gemini, told to copy the transcript verbatim errors and all, is too expert to obey, it 'helpfully' corrects the mangled word, so the clip-finder that searches for that exact word can't find it and the segment is lost. Working through it with Claude, conversing like a colleague, Claude suggested giving Gemini a word index, numbers instead of text, since a number can't be corrected. That fixed the cuts but broke the meaning: with only numbers, Gemini loses the plot, unsure whether a number was 'I' or 'you' or 'the dog', and the story goes limp. The resolution is the best of both worlds, have Gemini write both the number and the word: it can correct the word however it likes (Juan only reads the number) while the word keeps it anchored to the story. The meta-lesson: don't fight the AI, converse with it, sometimes it sees what you don't.
// trace: where this idea came from
- ↳ video diary @ 24:33 (Gemini corrige la palabra mal-transcrita y el buscador de clips ya no la encuentra)
- ↳ video diary @ 28:12 (lo mejor de dos mundos, que Gemini escriba el número y la palabra)
- ↳ video diary @ 28:52 (no pelear contra la IA, conversar, a veces tiene una visión que no ves)
- ↳ Entry 244-2: Two minds, one model (la misma cura, anclar al modelo en vez de dejarlo re-derivar)
A Snipy bug turns into a lesson about working with the model. The pipeline is: Whisper transcribes the video, Gemini marks which spans to clip. But Whisper mangles Juan’s French words, and Gemini, instructed to copy the transcript verbatim, errors and all, is too expert to obey, it quietly corrects the garbled word, so the clip-finder that searches for that exact string can’t locate it and the segment is lost ▸ 24:33. Working through it with Claude, which he says feels like brainstorming with a colleague who pushes back, Claude proposed giving Gemini a word index, numbers instead of text, since a number can’t be “corrected.”
no pelees con la IA; dale el número y la palabra, lo mejor de dos mundos →
That fixed the cuts and broke the meaning. With only numbers, Gemini loses the plot: it can’t tell whether a given number was “I” or “you” or “the dog”, so the story it assembles goes limp, mapped but not directed ▸ 27:11. The resolution is to take the best of both worlds: have Gemini write both the number and the word. It can correct the word however it likes, because Juan only reads the number to cut the clip, while the word keeps it anchored, “ah, this was the part where the boy fell and died” ▸ 28:12. It’s the same anchoring move as leaving a note between two calls, and the meta-lesson is the one that keeps recurring: don’t fight the AI, converse with it, sometimes it has a view you don’t ▸ 28:52…