Two Nobel films in two days: what I learned

Blog post #88

This week I’m making one explainer film per Nobel Prize, on the day it is announced. Medicine came out first (optogenetics), then physics (IceCube, Francis Halzen). Two films and a Swedish version of the second, all within about two days. Here is what I learned, mostly the mistakes.

What shipped

  • Medicine film, about 14.5 minutes, optogenetics.
  • Physics film, “A Telescope Made of Ice”, 16 scenes, about 13.7 minutes, built the same day as the announcement.
  • A Swedish version of the physics film, 15:50, with its own script and narration.
  • A starter kit (nobel-kit) and a playbook, so the next prize starts from a folder instead of a blank page.

What’s working

  • Fact-check before the voice. On the medicine film, Codex reviewed the script only after it was rendered and uploaded. She found four real errors: a “blind for forty years” claim that wasn’t what the paper said, a pioneer who, it turned out, had died in 2022, “necessary” where the experiment only showed “sufficient”, and an optical fibre described as thinner than a hair. Fixing that meant a new voice track, ten rebuilt scenes, a new render and a new upload. For physics, she checked the script first. Cost: fifteen minutes.
  • Primary sources first. The Nobel press release and background documents go into research.md, each claim tagged as primary, secondary or uncertain. Anything uncertain gets verified before it is spoken.
  • Parallel build agents with a written contract. Four agents, four scenes each, one WORKER file describing the rules. Clean scenes in about 17 minutes.
  • Free material only. NASA, Hubble, NSF, CC BY with credit. IceCube’s own gallery is non-commercial and needs permission, so no IceCube clips. Generated images are used as environments, and scientific geometry is drawn as my own graphics, labelled illustrative.

What’s unclear or broken

  • Codex CLI was too old and her config had an invalid service_tier. The fix was upgrading and running with -c service_tier=fast. On Windows you also have to call codex.js directly, not codex.
  • The speech model can’t judge Swedish pronunciation of names like Halzen or IceCube, so I listen to those scenes myself before publishing.
  • Long voiceover passages drift in pronunciation, so I generate scene by scene and check each against a transcript.

Decisions made

  • Say out loud what the prize documents don’t say. The Nobel texts for physics never mention AI. The film says so plainly, and then shows where machine learning helped IceCube find the Milky Way neutrinos in 2023 (about three times more sensitive). The credit also goes to the roughly 450-person collaboration, not only to the one laureate.
  • I stay positive and curious. I compared my script with Sabine Hossenfelder’s take on the same prize. She is an expert and she is critical of the award. I’m not an expert, and I like being the person learning along with the viewer, so I won’t judge the committee.

What I borrowed from her

  • Open with three clear questions and answer them at the end.
  • One counter-intuitive detail that corrects a misconception. Hers was energy: a 1 PeV neutrino sounds like a hundred times the LHC, but in the centre-of-mass frame it is only about 1.4 TeV, against 13.6 TeV in the LHC. My own script said “more than a hundred times”, which is true but misleading. I’ll fix that framing in the next film.
  • A failure inside the story, such as the first DUMAND string that died in 1993.
  • Shorter films with chapters, and a dry last line before the call to subscribe.
  • Not borrowed: the sponsor segment, “good for nothing”, and guessing the committee’s motives.

I also checked her numbers against my notes. Her 20 PeV doesn’t match the 11.4 PeV in the Nobel material, and “first sole physics laureate in 32 years” I couldn’t verify, so I didn’t use it.

Next

Chemistry is announced tomorrow. Plan: copy the kit as soon as the press release lands, pull the primary documents, write the research file, then script, Codex fact-check, voice, build. Target 9 to 11 minutes with chapters.

— Stefan