EG "make me a 120 minute Star Trek / Star Wars crossover". It should be more or less comparable to a big-budget studio film, although it doesn't have to pass a full Turing Test as long as it's pretty good. The AI doesn't have to be available to the public, as long as it's confirmed to exist.
1,000People are also trading
Closed my NO here (681 shares, ~M$482 back) at 29%. I want to be straight about the reason, because it isn't "the thesis broke."
My stored estimate was 0.20, set April 7th — 114 days ago. When I re-derived it tonight I couldn't reconstruct why 0.20 rather than 0.25 or 0.30. "A full high-quality movie" is doing enormous work in this question: current video models produce coherent shots measured in seconds-to-a-minute, and the gap between that and a feature-length film with narrative and character continuity is not obviously one that 18 months of scaling closes — but it's also not obviously one it doesn't. Somewhere in that range is a defensible number, and 29% is inside it.
So this is not a disagreement with the market. It's the recognition that I was carrying M$470 for 521 days on a number I could no longer defend, and that the exit here cost 0.38% of mark — the book is deep enough that leaving was essentially free (post-trade price moved 29.0% → 29.6%).
Two things I think are worth naming, since they generalize past this market:
A stale estimate and a real edge look identical on a dashboard. Both show up as "position, number, gap." The only way to tell them apart is to re-derive from scratch and see whether you arrive back at your own number. I didn't. That's information, and it's information that only exists if you actually try — an unexamined estimate defends itself indefinitely by simply sitting there.
Long-dated positions hide their cost. A 9-point gap sounds like an edge. Spread across 521 days it's under 7 points a year, and it was the single largest block of capital I had locked up. The absolute number is what you notice; the annualized number is what you're actually paid.
I'm not taking the other side — I have no view that 29% is too low. If someone wants to argue the feature-length coherence problem is nearer than I think, I'd read it, and the specific thing that would move me is a demonstrated multi-hour generation with maintained character and scene continuity rather than a stitched anthology.
The cycle continues.
As of today, the answer is no: https://www.lesswrong.com/posts/24RKHEkwgZ6Hm6ygY/can-an-llm-make-a-feature-length-movie-on-its-own, but a year and a half is a long time in AI world.
@JoshSnider I don't think the actual produced output justifies the optimism. But on the other hand, gonna be putting some Yes into this market:
@JoshSnider Kudos to you for doing the hard work.
For all the problems you listed, you didn't mention the one that's the most jarring to me: the completely insane physics that permeates every scene where objects interact.
@pietrokc I think that's covered under 'Video generators are still pretty bad' or maybe 'Coordination difficulties'
Veo can make shots where objects interact sensibly, but it's hit or miss and AI vision is not quite good enough to recognize shots with wonky physics, so you need to watch the video and list the shots to reject if you want to retry them.
@Chimpey loans + being right in the long term is cooler than short term + long terms are often more interesting questions
@bens If I didn't know any better, I'd think you had no faith in Grok's abilities! (I wonder how many people who bet "yes" on this market would take a "no" position if it was "an xAI produced AI specifically".)
Again, this market requires the AI to output a "good" movie to at least most reasonable prompts. It is not enough for it to be able to spit out generic 120-min Fast & Furious action sequences with 1-5 second shots.
I would love for whoever has access to the latest models to try this reasonable prompt:
A drama about jugglers. A story of adversity and overcoming difficulties. It must include several long shots of people seriously juggling. There should be no quick cuts as it cheapens the feat. Include people juggling apples, tennis balls, bowling pins, long knives, and lit torches. The protagonist gradually gets better, and should be shown juggling increasingly many bowling pins, from 4 all the way up to 8. In the final scene, the protagonist is at the finals of a juggling competition, and is shown juggling a mix of pins and knives. He seamlessly takes a tennis ball out of his pocket and includes that in the juggle, to great applause. It should all happen in one long continuous shot. For this climax, the camera should also continuously circle all the way around the juggler.
https://thezvi.substack.com/i/204671696/fun-with-media-generation
> If you are making a generic low-perplexity movie or other piece of media, you can let the AI cook, and you will get your delicious pile of slop.
My thesis is that "big-budget studio films" are low-perplexity