From 13:00 announcement to a two-language film the same evening
Blog post #92

AI-generated illustration.
The Nobel Prize in Literature was announced at 13:00 today. It went to Anne Carson, the Canadian poet and classicist. By about 19:30 the film was public on YouTube, in English with a Swedish audio track on the same video: Anne Carson Wins the 2026 Nobel Prize in Literature.
Literature was the prize I worried about most. There is no experiment to animate, no data, and at first glance no link to AI. It turned out to be one of the most fun films to make.
What shipped
- A 9:20 film in English, with a Swedish audio track, Swedish title, Swedish description and a Swedish thumbnail on the same upload.
- An intro I built before anyone knew the winner: a real photo of the Swedish Academy’s building at dusk, light coming on in the two real doors, letters flying out of them, and a placeholder name. After 13:00 I only had to fill in “Anne Carson”, her book titles and Greek letters.
- Paintings made with AI where the ancient world meets ours: the oracle at Delphi checking prediction-market odds on a phone, satyrs dancing a tango, a red winged boy with a camera (Geryon, from her best-known book), a dove over Oslo for tomorrow’s Peace Prize.
- Real material next to them: a Sappho papyrus, a 14th-century manuscript of the Catullus poem she uses in Nox, an ancient vase with Herakles and Geryon, portraits of Woolf, Brontë and Keats, and photos of Carson herself under a free licence.
What’s working
- Preparing before the news. The intro and outro were done at 12:50, so the afternoon went to the script and the scenes.
- Several reviewers. One AI model fact-checked every version of the script and found real errors: the wrong title for her thesis, a prize year off by one, claims about the Sappho brackets that were too strong. ChatGPT read the Swedish and caught phrases that sounded translated.
- One file, two languages. The script now holds English and Swedish side by side for every scene. The same file produces both voices, each scene is as long as the longer language, and a small tool mixes the Swedish voice into one audio file exactly as long as the video. That file becomes the Swedish track on YouTube.
- Four agents, then a patch round. Four parallel agents built the scenes. After my review they patched their own scenes from a change list instead of starting over, and I re-rendered only the one scene that changed and spliced it in.
What I changed in the script
I rewrote it six times, mostly because of my own reading:
- Start with the news. The first line is now the winner, then who I am and what the film covers.
- No clever frame. I tried an Odyssey frame for the whole film. It made the text harder to follow for anyone who hasn’t read the Odyssey, so it went.
- No sources in the voice-over. “On Kalshi”, “on Swedish television, the critic…” broke the flow. Sources now sit in small glass labels in the picture.
- Paraphrase rather than quote. Except for the prize citation, one Sappho fragment and one line of Latin, which are there so you hear her language.
- No short, chopped sentences. “So, where is the AI? The Academy doesn’t mention it. But remember the brackets.” That rhythm sounds like AI writing. Full sentences now.
What’s unclear or broken
- The YouTube API says yes and no to the same token. The upload worked, then everything after it failed with 401 “Invalid Credentials”. I renewed the login and gave the scripts a shared client that logs every call and retries. Then I tested it properly: the same access token went 401, 401, 200, 401, 200 within eight seconds. Retrying gets through, but the cause is somewhere on Google’s side or in how the app is set up. The log will show if it gets worse.
- Pauses. The Swedish voice was longer than the English one in every scene, so the English version had long silences while the picture waited. I shortened the Swedish text instead of adding pauses.
Decisions made
- Thumbnails follow the series look: the same lavender background, the person’s name, “Nobel Prize 2026”, and three words about her (poet, classicist, translator) instead of tags like “explained + AI”.
- No “Explained” in the title.
- Swedish is a second audio track, not a second upload.
Tooling & process
- The secrets the scripts need are kept in a secure vault, never in a file. Today the scripts got permission to write a renewed login token there themselves, so the next renewal doesn’t need me.
- Tomorrow at 11:00: the Peace Prize.
— Stefan