My friend Jay and I catch up every month or so on Thursday night video calls, which is Friday morning where he is, in Australia. This week Jay said the timing is convenient because his weekly Claude Code allowance resets on Friday afternoons, so Friday morning is a stretch of his week when he’s out of tokens and free to talk.
Aswin, a former Sitecore colleague of mine, recently joked, “Life is what happens while you wait for your AI compute to reset.”
Early adopters of frontier AI did not set out to live in a world where the hardest thinking gets scheduled around when a meter refills. A year ago there wasn’t even a usage meter.
Token usage windows now have the capacity to shape your week.
Jay and I talked through recent experiments and prototypes, mine for touring artists and his recent research into feeding unstructured data into processing systems that generate video. He told me about HeyGen’s new open-source framework that turns HTML into video called HyperFrames. Coming out of our call I started experimenting, and my first project focuses on a real estate listing, tags every photo into a fixed set of room categories, and writes a narrated script where every spoken claim traces back to the listing description, a spec field, or something visible in the photo on screen.
Through your AI app of choice, HyperFrames enables backend video generation (applying transitions or Ken Burns effects, adding chapter titles, animating graphics, even aligning a voiceover script to photos with music underneath).
I hooked Claude Code up to HyperFrames over MCP, and 10 video runs produced 28 minutes and 24 seconds of finished video in 20 minutes and 52 seconds of render. So that’s about 44sec/min, seconds of render per minute of finished video. The real-time break-even line is 60sec/min, so anything under it renders faster than you can watch it.
| Output | Runs | Avg Render | Render Rate |
| 720p | 5 | 90 sec | 31sec/min |
| 720p + titles | 3 | 157 sec | 54sec/min |
| 1080p + titles | 2 | 164 sec | 65sec/min |
Motion and audio alone come in at 31sec/min, about half of real time. Titles and counting graphics are what push it past break-even.
These are still early days, and I think I’ll be incorporating this into my video editing workflow for things like animations, text and captions.
My quick test videos are up on YouTube, the first listing and the second. They’re basic proofs of concept. I almost never see real estate video on YouTube, which makes me think the opportunity is scaling this to generate a real estate agent’s entire active portfolio of properties at once. If you sell houses please reach out, because I’d love to know if you’ve seen a solution like this before.
Now, let’s put the two facts together.
- Production quality video was prompted and generated without a track-based timeline like Final Cut or Premiere.
- AI video renders faster than you can watch it.
I think that leads, in a few years, to a stream of content that stops being a library you choose from and becomes a feed that’s custom-rendered for whoever is sitting in front of it.
Netflix ran this live last year as a joke. “Bête Noire” is a Black Mirror episode about a woman whose memories stop matching everyone else’s, and at one point the characters argue over the name of a fried chicken chain. Netflix shipped two versions of the episode and handed them out at random. (A) In your version, the cap in the opening scene reads Barnie’s. (B) In your neighbor’s version, the same cap reads Bernie’s. Your version stayed locked to your account, so signing in on another device wouldn’t show you the other one. The plumbing to deliver different versions of a TV episode was already boring enough to use for a gag.
What’s new this year are capable consumer-level AI video content generation engines.
Imagine a near future where we’re watching television episodes with the same script, same actors, same arc, with a pace that’s personalized for the viewer. So, the same episode might be a little more like a thriller for you and a little more like a rom-com for your spouse. The music would be a slightly different score, and nobody has to select the songs because they can be chosen for us. I don’t think I want to watch that show.
I used to host Breaking Bad watch parties at my rental house in the Little Five Points neighborhood of Atlanta. The show was excellent, and during the week between episodes we argued about what Walt would do next. Everyone was wrong in a different way, and the next Sunday settled it for all of us at once. ABC’s show Lost fueled similar fan theory energy before losing its path, and companion podcasts like the one Ben Stiller and Adam Scott release for Severance keep that energy alive, even as they hand you the theories instead of leaving you to build your own.
The wait is mostly gone already, traded away for a full season at midnight. The episode survived that trade, because we all still got the same episode. Personalized generation takes the last piece. If your version was paced for you, I can’t argue with you about it.
If we’re no longer watching the same thing, there’s nothing left to wonder about together.
This was orignially featured in The Catalog newsletter
Leave a Reply