The talk is itself an AI game: runtime decision-making running live, and an NPC match that plays out differently every time
Think You Can Build a Game with AI? Think Again! - Danielle An & David Hoe, Meta · David Hoe
18 min total·Actually worth watching closely: ~4 min·3 must-watch clips
- 0:15 – 1:30Watch
Opening: you are playing this talk
The speakers ask the room to scan a QR code and steer the character on stage with their phones to reveal the slide content; when several people grab control at once, a voting system decides whose input wins. The medium of the talk is itself an AI-generated game.
They are not going to tell you what AI games can do — they have you play one first, then draw the conclusions.
The audience collectively steering the character happens only on screen; on audio alone you get scattered laughter and stage patter and no sense of what is actually going on.▶ Jump to 0:15Speaker · David Hoe - 1:30 – 4:10Listen
The barriers are gone — now what
A coder who cannot draw and an artist who cannot code can each now take a game end to end alone. Then the speakers pour cold water on it: once everybody can build a game over the weekend, a couple of prompts get you something impressive for a moment, and then you are asking what's next.
What the tools flatten is the skill barrier, not how distinctive your work is — the immediate consequence of barriers falling is sameness.
This stretch is the two of them talking in turns, with the visuals purely decorative; listening while commuting or doing something else is enough.▶ Jump to 1:30Speaker · David Hoe - 4:10 – 7:18Skim
A single key art image as anchor
The speakers walk through a workflow their team actually uses: settle on one key art image first, then generate every asset against it so the style holds together on its own; more interestingly, that image often anchors what the gameplay could be as well.
Lock one image before you fan out — currently the cheapest defense against AI-generated assets that all talk past each other.
The visuals are a workflow diagram and an asset comparison; one glance at the anchor image and what it produced is enough, no need to watch frame by frame.▶ Jump to 4:10Speaker · Danielle An - 7:18 – 8:48Listen
When polish is table stakes, taste decides
The speakers list what still creates the gap: aesthetic quality, cohesion between UI, story and art living in one world, and a lot of play tests with real people. Once the tools flatten all of that, what is left is taste — the feel for which subset of humans would like this game and what they are really pursuing.
Taste is not an aesthetic adjective, it is concrete judgment about what your target players want, and that is the piece the tools cannot replace.
The whole segment is argument, with the screen parked on text bullets; the speakers' phrasing carries far more than the slide does.▶ Jump to 7:18Speaker · Danielle An - 8:48 – 10:22Watch
The unscripted multiplayer demo
A demo the team built in a couple of days: players assign each character a personality, then the characters are entirely LLM-driven in real time, deciding on their own to steal cubes, block, or kick others — nothing scripted, no two games alike. The speakers stress that 18 months ago the inference speed and cost made this simply impossible.
Once inference is fast and cheap enough, what appears is not a better version of old games but a genre that could not be made before.
The few seconds where characters change their minds and outmaneuver each other are the most convincing evidence in the talk, and you can only get them by watching; put into words it degrades into an ordinary feature description.▶ Jump to 8:48Speaker · Danielle An - 10:32 – 13:46Listen
From waterfall pipeline to parallel
Traditional development is an expensive straight line from design to art to modeling to animation to coding, where one stuck link stalls everything downstream. The speakers say that line is now broken up into parallel work, with an iteration compressed from months into hours or days.
The time saved does not go into making more content — it goes entirely into play tests and polish, and that is why the games get more fun.
It is a process comparison with two pipeline diagrams on screen; catching the line about spending the saved time on play tests matters more than looking at the graphic.▶ Jump to 10:32Speaker · David Hoe - 13:50 – 16:15Listen
The whole stack is nondeterministic now
On a platform where everybody creates, every link is model-driven: the prompt the player writes, the runtime model making decisions, and how content gets ranked and delivered. The speakers say plainly that this breaks the engineering habit of writing tests for stability and debuggability — a model upgrade or one changed prompt can flip the system's behavior entirely.
With undeterminism throughout the stack, debugging and operations have no mature answer yet — a problem the speakers admit they have not solved.
This is the most candid self-critical stretch of the talk; the value is entirely in the speakers' wording and tone, and the screen adds nothing.▶ Jump to 13:50Speaker · David Hoe - 16:15 – 17:40Watch
The reskin demo, and what you can do this weekend
First the whole game world is turned into the Labubu universe on stage, showing how the model can adjust the experience for one individual player in real time. Then come layered suggestions: if you have never built a game, use any model you like this weekend to prompt something Tetris-sized and find out firsthand where it works and where it fails; if you are past that, upgrade it with art, UI and surprises so it does not feel prompted out of the gate. It closes on two unanswered questions — how creators, players and the platform all stay profitable under a token economy, and how you keep runtime-generated content safe.
Building one bad game yourself beats reading ten more analyses; as for money and safety, the speakers have not cleared those two either.
The world reskin in the first half is a live demo you need to see; the advice and open questions after it are fine to just listen to — watch through the demo, then switch to audio.▶ Jump to 16:15Speaker · Danielle An