Demo porn – one-shot AI games are not games
- https://www.youtube.com/watch?v=LmXU6SEH3Ks
- Original title: Stop Being Tricked
Prime pushes back on the wave of Claude Opus 5 "one-shot game" demos circulating online — Call of Duty clones, Starfield-alikes, snowboarding games generated from a single prompt. He refuses both the "it's fake and sucks" camp and the "game devs are unemployed" camp. The demos are genuinely impressive as a capability milestone, but they produce a point in the space of possible games, never the specific game someone wanted, and the gap between the two is where all the actual work lives.
The crab argument
Ask an image model for a crab and you get something crab-shaped, and it looks amazing — because you had no opinion about what the crab should be. Any member of the set of all crabs satisfies the request. The moment you have a specific crab in mind (exact lighting angle, exact motion, exact details), the distance between "a point in the space" and "the thing I want" becomes enormous. One-shot game demos exploit the same trick: the prompt has no opinions, so anything plausible reads as success.
Minimal prompting is the tell
Matt Schumer, driving many of these one-shot demos, wrote up his method, and the key instruction is that minimal prompting is better — let the agent decide the specifics. Prime reads that as the admission: you can get anything you dream up, just nothing you actually want.
His own attempt makes the point concretely. A stick-figure golf-swing game generated in minutes, but the figure looked wrong, the swing was weird, the character faced the ball incorrectly. Two hours of iteration, $117 of Claude Opus 5, 72 million tokens — and the result was roughly 25% of what he had envisioned.
Demo to shipped game is the whole job
A visual clone of a shooter is not a shooter. What makes a multiplayer shooter good is level design, and a single good multiplayer level takes hundreds to thousands of hours of precise iteration. Follow any of these demos toward an actual product and the discoveries stack up: hiring an artist is 10x faster and better for modeling, performance problems need an expert, fun level design needs someone who understands game mechanics. You end up with the same team you always needed, having generated something that merely looked compelling.
The common response — "anyone who knows Unity could do that in two hours without spending a fortune" — misses the point in Prime's view. That the generation is possible at all is cool. The point is that generating exactly what you want is still unsolved.
AI did not solve the abandoned-project problem
If anything there will be an order of magnitude more abandoned projects. Even with full generation, the tedium and back-and-forth stay; you move faster into a new set of problems. Good decisions do not come from "model, you decide" — they come from sitting with one small feature for days. Prime's example from a tower-defense game he is building: a bullet-time effect triggered on upgrade took four or five hours of experimenting to get the feel right, and a game is an amalgamation of a million such features.
Quoted framing
He replays a clip making the same argument from the other direction: if the technology let you clone GTA in three years instead of 30 seconds, people would happily spend the three years — and that technology already existed before AI. But it still would not be GTA, only a clone. Hits are unexpected by nature; cloning cannot produce them.
Takeaway
Not "so over", not "so back". Don't be tricked by demo porn. In a world where generation is free, determining what is actually good remains the hard, slow, human part.