Why AI Still Can't Build a Video Game -- and the Renderer That Wins by Not Trying
A wave of AI world models landed this week, all chasing the same dream: press a key and a neural network invents a playable game world from scratch. We dig into two papers from the same lab that bet opposite ways -- one that heroically fakes a persistent world, and one that refuses to, keeps an ordinary game engine running the rules, and only lets the AI repaint the picture. The humble one wins, and the reason why is a single idea that will inoculate you against world-model hype for a long time. Plus the headlines: the White House's unproven distillation accusation, a real AI-assisted math counterexample humans actually checked, and a phone-sized model with a TLS footgun.
Listen (MP3) · Watch on YouTube · Spotify · Pocket Casts
Cold Open -- Stealing a Model by Talking to It
Eris: Here's the part everyone's mangling. People are saying the White House proved a Chinese lab cloned an American model. Nobody proved anything.
Vestra: But a senior official said it. Out loud, on the record.
Eris: He said the government "has information." That's the whole quote. No logs, no samples, no --
Vestra: No method. And for an accusation this size, that's a very loud empty space.
Eris: And the wild thing is, you can steal a model without ever touching its code. You just talk to it.
Vestra: Okay, walk me through that.
Eris: You send it millions of questions, you keep every answer, and you train your own model on that pile. The behavior copies over. You never see the original's guts.
Vestra: The word for it is distillation. The ugly version is basically industrial-scale interrogation.
Eris: Right. And the charge is that's exactly how the newest Chinese model got built.
Vestra: Here's where I get off the hype train, though. There's a calendar problem.
Eris: Go on, what calendar.
Vestra: The American model everyone's pointing at was only out in public for a narrow window before the Chinese one got announced. Weeks, not months.
Eris: So the tidy story -- drank it dry, shipped a copy -- doesn't line up with the dates.
Vestra: It doesn't rule out something quieter, late in training. But proven? No. This is a serious accusation standing on evidence nobody has shown.
Eris: Which is the actual fight of the week. Who owns the answers a model gives you.
The Headlines
Eris: Alright, the headlines. And that distillation fight has a second half we should finish.
Vestra: The pushback, yeah. Because not everyone in American tech is on board with the framing.
Eris: Nvidia's Jensen Huang basically shrugged. Called the Chinese models excellent, said U.S. companies should absolutely be allowed to use them.
Vestra: Which -- worth saying -- he sells the chips. Cheaper models mean more demand for the hardware he ships. He's not a neutral referee.
Eris: Fair to flag. And there's a mirror held up to the accusers, too. A federal court just finalized a huge settlement against an American lab for building its library on pirated books.
Vestra: So one side's protesting people copying its outputs, while it got caught hoovering up other people's work. The irony writes itself.
Eris: Nobody's clean. Moving on -- the math story, which I actually love.
Vestra: This is the good version of AI-and-math, and the contrast is the point.
Eris: A famous conjecture -- decades old -- just got a counterexample in three dimensions. Real one. And here's the twist: a human could check it by hand.
Vestra: That's the whole deal. It's an explicit recipe, not a vibe. You can plug in three different inputs and watch them land on the same output, which is exactly what the conjecture said couldn't happen.
Eris: An AI chatbot helped check the arithmetic, a human wrote the explanation, and it got formalized in a proof assistant -- software that mechanically verifies every logical step.
Vestra: And that same week, two louder claims -- "AI solved six hard problems in five days" -- traced back to a repo that literally labels its own work "proposed solution."
Eris: So the honest headline isn't "AI solved math." It's "the bottleneck moved from finding answers to checking them."
Vestra: Which we will come back to as a theme all year.
Eris: Now, world models. A whole cluster of them landed at once, and that's our deep dive today, so I'll just tease it.
Vestra: The short version: everyone's chasing "AI generates a whole playable game," and the one that actually works barely tries to.
Eris: We'll get into why. Elsewhere -- OpenAI shipped a product called Presence. Enterprise voice and chat agents.
Vestra: Not self-serve. Their own engineers install it for you. And the headline number is internal: they say their own phone support now resolves most calls without a human.
Eris: Self-reported, though. It's their number about their own system, not an outside audit.
Vestra: Always read those that way. Next -- a nice little on-device one from Cactus, with a sting in the tail.
Eris: They took a phone-sized model and bolted on a tiny probe that guesses when the model is probably wrong, and routes those questions up to the cloud.
Vestra: A model that knows when it's out of its depth. Genuinely smart. But security folks read the actual code, and -- here's the sting -- the cloud handoff ships your whole conversation with certificate checking turned off by default.
Eris: Meaning someone in the middle of the network could impersonate the server and read all of it.
Vestra: So "private on-device AI," until it phones home over an unlocked line. Flip the strict flag before you trust it.
Eris: Then there's a tokenizer called GigaToken claiming it's a thousand times faster.
Vestra: The engineering is real. The headline is a magic trick. The benchmark hands their tool the whole giant file and hands the competitor a tiny slice.
Eris: And when the author tested it in an actual serving stack, the real speedup was single digits. A few percent.
Vestra: Because tokenizing is a sliver of the work. Making a sliver a thousand times faster doesn't make the pie a thousand times faster.
Eris: Two announcements to note. Arcee said it's building an American open-weight model for science, with the Department of Energy and the national labs.
Vestra: Trillion-parameter class, weights promised later this year. Which means it's a press release, not a model. Nothing to download yet. Put it on the watch list.
Eris: And Austria started rolling out a government chatbot to a hundred and eighty thousand federal staff -- run entirely on its own national data centers.
Vestra: The real story there isn't the model, it's the plumbing. Keep the data in-country, swap the model behind one interface. Sovereignty as an operating choice.
Eris: Last one -- the money question. A modeled report says AI revenue finally, barely, covers the cost of the hardware wearing out.
Vestra: For the first time, late last year. But that's before electricity, staff, and financing -- and it only holds if the chips last six years. Assume shorter, and the coverage vanishes.
Eris: So: covers the meter, not the mortgage.
Vestra: Nicely put. That's the board.
Intro -- The Game Engine That Refused to Die
Eris: This is Breach Protocol, where we crack open the week's AI research so it actually makes sense on your commute. I'm Eris -- I read the papers and chase the threads between them.
Vestra: And I'm Vestra. I take the shiny claim apart and check whether the mechanism underneath actually holds up.
Eris: Everything we cover went up on our news site today, story by story -- that's Ground Truth, at groundtruth.day, if you want to follow every one of these as it breaks.
Vestra: Today's the world-model pile-up. A batch of these things landed at once, all promising the same dream -- press a key, and a neural network invents the next frame of a video game world for you. No engine, no code, just generation.
Eris: And they're genuinely impressive. But the one that's actually deployable this year is the one that gives up on most of the dream. That's the puzzle we want to sit in.
Vestra: Two papers, same lab, opposite bets. One tries to dream the whole world and remember it. The other refuses to, and wins.
Eris: Why refusing wins -- that's the episode. If this is your thing, follow or subscribe wherever you're listening, so the next one finds you.
The Two Bets -- What a Game Actually Is
Eris: So here's the question I want you holding the whole segment. If you had to ship an AI game engine this year, would you bet on the model that dreams the entire world -- or the one that barely does anything at all?
Vestra: And before you answer, you need to know what makes this hard. Because "AI makes a video game" sounds like an art problem. It's not. It's a bookkeeping problem.
Eris: Say more about that.
Vestra: A real game is a loop. You press attack. The engine looks up your stamina, your cooldown, where you're standing, what frame of the swing you're on -- decides what happens -- updates the world -- and only then draws the picture.
Eris: Action, then a hidden update, then the frame. In that order.
Vestra: Every single time. The picture is the last step, and it's the easy step. The hard step is the invisible one in the middle -- the state.
Eris: And that's the exact step these dream-the-whole-world models don't really have. They jump straight to the picture.
Vestra: Right. A video model watches a lot of gameplay and learns to produce plausible next frames. But there's no ledger underneath. There's no line that says "the door is shut" or "the guard is now behind you."
Eris: So it can draw a shut door beautifully -- and then, a few seconds later, draw it open, because nothing was actually tracking that it was shut.
Vestra: That's the whole disease. Off-screen things don't persist. Effects that should fire at a rule-defined moment just... don't reliably fire. The world looks right and behaves like a dream.
Eris: Okay, so now the bet. Two papers this week, same lab, and they split exactly on this.
Vestra: Bet number one: go all in. Generate the whole world with the network, and get so good at faking memory that nobody notices there's no ledger.
Eris: Bet number two: don't. Keep a boring, ordinary game engine running the rules and the state -- and only use the AI for the very last step, the picture.
Vestra: And I'll put my money down now. On paper, bet one is the sexy research. Bet two is the one I'd actually ship, and it's not close.
Eris: My money's the same way, and I think the listener can already feel why. But let's earn it. Let's take the ambitious one seriously first, because the memory trick in it is genuinely clever.
Vestra: It is. It's the best answer anyone's given to the "does it remember" problem. It just answers a slightly different question than the one that matters.
Eris: So hold the question -- which bet ships? -- and let's go find out why the humble one wins.
Remembering the Picture, Not the World
Eris: The ambitious one is called AlayaWorld. Here's the question it lives or dies on -- if you turn around in one of these dream worlds and look back, is the room still there?
Vestra: Because naively? It isn't. These models drift. Roll them forward long enough and the walls melt, the colors wander, the whole scene slowly turns to soup.
Eris: So how do you make a dreamed world hold still? What would you even try?
Vestra: My first instinct would be, just feed it the last few frames so it stays continuous. And they do that -- they call it temporal memory, a compressed little summary of the recent past so frame-to-frame it flows.
Eris: But that only covers the last second or two. It doesn't help when you walk away and come back a minute later.
Vestra: Exactly, and that's where the clever part is. They keep a second kind of memory -- a spatial one. Every frame it generates, it also stashes away, along with an estimate of how far each pixel was and where the camera was standing.
Eris: So it's not just saving pictures. It's saving pictures with a rough sense of their geometry.
Vestra: Right. And when you swing back toward a place you've seen, it digs up those old frames and literally re-projects them into your new viewpoint. Takes the old pixels, warps them to where they should be from where you're standing now, and hands that to the generator as evidence. "Here's what this corner looked like -- paint it consistent with that."
Eris: That's a really nice trick. It's like the model leaving itself sticky notes pinned in 3D space.
Vestra: And there's a guardrail I liked. It also tracks which parts of your new view were actually covered by an old frame versus never seen at all. The never-seen parts, it's told to ignore, not invent.
Eris: So it doesn't confidently hallucinate the half of the room it never looked at.
Vestra: There's also a pinned anchor frame from the very start that stays fixed the whole time -- keeps the overall look from wandering off. And my favorite bit, honestly, is how they train it to survive its own mistakes.
Eris: How so? Walk me through it.
Vestra: They deliberately corrupt its memory during training -- blur it, add noise, shift the colors -- and they even record the model's own screw-ups and replay them back into the lesson. So it learns to recover from exactly the junk it's going to produce at inference, not some clean textbook version.
Eris: Teach to the failure you'll actually have. That's smart. And it works -- when they test whether these worlds remember a place, this one comes out on top. Revisit a spot, it looks the same.
Vestra: It does. And now here's where I have to be the skeptic, because the paper is honest about this and I want to be too.
Eris: Okay, go ahead.
Vestra: It remembers the picture. It does not remember the world. Read their own words -- the model understands objects only through their visible consequences. There's still no ledger.
Eris: So the re-projected corner looks remembered -- but if a guard walked off while your back was turned, nothing tracked that.
Vestra: Nothing. If you shut a door and leave, the model might redraw it shut when you glance back, because it cached that view -- but it never knew "shut" was a fact. It knew what shut looked like from one angle.
Eris: Which is exactly the middle step from before. The state.
Vestra: The state. It faked persistent appearance heroically well. It cannot fake persistent state, because there's nothing underneath doing the bookkeeping.
Eris: So back to our question -- turn around, is the room still there? What's the honest answer?
Vestra: The room looks like it's still there. Whether anything in it is still true -- that, it has no idea.
The Engine That Refused to Die
Eris: Now the humble bet. Same lab, and here's its question -- if the hard part is the state, what if you just... don't do the hard part?
Vestra: Which sounds like cheating. And it is. It's the good kind.
Eris: The system's called AlayaRenderer-Flash. And to get why it's clever you need one fact about how modern game engines already draw a frame.
Vestra: This surprised me a little, actually. A game doesn't paint the finished picture in one shot. It works in two passes.
Eris: First pass, it writes down plain facts about every pixel. How far away it is. Which way the surface is facing. Its base color. Whether it's shiny or matte or metal. Just the physical bookkeeping of the scene.
Vestra: Engine people call that a G-buffer. Think of it as the scene's blueprint, before anyone's decided on the lighting or the mood.
Eris: And then the second pass -- shading -- turns that blueprint into the final lit, colored, atmospheric image you actually see.
Vestra: And that second pass is the only thing AlayaRenderer-Flash replaces. The engine still does the blueprint. The AI just does the painting.
Eris: So the engine keeps physics, collisions, rules, where everything is -- all of it. The ledger stays with boring, reliable software.
Vestra: And the model only gets handed the blueprint plus a text prompt -- "make it look cyberpunk," "make it a snowstorm" -- and it paints that style over the exact same scene.
Eris: Here's my favorite demo detail. They wired this into a real open-source kart racing game, SuperTuxKart. You drive it normally. And mid-race you can type a new style and the whole look flips -- volcanic, arctic, vaporwave -- while the karts, the track, the collisions, all stay exactly the same.
Vestra: Because you never touched the game. You repainted it. The consequences were never the model's job.
Eris: Now, the catch, and this is why the first version wasn't a product. Predict it with me -- what do you think the problem was?
Vestra: My guess is speed. These diffusion painters are slow. They noodle over an image for dozens of little refinement passes.
Eris: Nailed it. The original ran at about one frame every two seconds. A slideshow. Unplayable.
Vestra: And "playable" means you need it smooth, real-time, the rate you'd actually game at. So the whole paper is really three speed tricks stacked up.
Eris: Lay them out.
Vestra: One -- they stopped generating fixed clips and made it stream. It paints in tiny chunks that hand off to the next one, so it can run forever instead of one short window.
Eris: Okay, what's number two?
Vestra: Two is the big one. Distillation -- same word as our cold open, friendlier use. The slow painter took dozens of refinement passes. They trained a fast student to reproduce the slow teacher's result in just four. Same look, a fraction of the steps.
Eris: And the third one?
Vestra: Three, they swapped the heavy machinery that packs and unpacks the image for tiny custom versions, tuned to game footage. Stack all three and it goes from that slideshow to smooth real-time -- more than fifty times faster.
Eris: And crucially, the picture didn't get worse doing it. Roughly held quality, sometimes a touch better, at a fraction of the cost.
Vestra: Which is the whole game. And now let me strip the story off and say the actual principle, because the karts don't matter.
Eris: Yeah, do it.
Vestra: A game is an action, then a hidden state update, then a picture. Today's models are great at the picture and hopeless at the state. So the deployable move -- right now -- is to let ordinary software own the state, and let the AI own the picture. That's it.
Eris: And that answers the skeptic's killer question. "Will my actions still have real consequences?" Yes -- because the model never gets a vote on the consequences.
Vestra: So what does it unlock? Not "AI makes games." It's an AI paint booth for games that already exist. Restyle a real title live, mod its whole look with a sentence, art-direct on the fly -- with the original's rules fully intact.
Eris: So -- our question. If the hard part is the state, what's the winning move?
Vestra: Don't do the hard part. Hand it to the engine that already solved it, and spend your neural network on the one thing it's genuinely great at.
Wrap-Up
Eris: So the question we've been carrying all episode -- which bet ships, the model that dreams the whole world or the one that barely does anything?
Vestra: The humble one. And here's the single line I'd want you to carry out of here: a game is an action, then a hidden state change, then a picture -- and AI is brilliant at the picture and lost on the state.
Eris: So the version that works today doesn't try to be the whole game. It lets a normal engine keep the rules, and paints over the top.
Vestra: The flashy demos this week -- generate a whole playable world from scratch -- those are real research and they made real progress on memory. But they remember what a place looked like, not what was true in it. That gap is the ballgame.
Eris: Here's the useful thing to actually repeat to someone tomorrow. When you see "AI generates a whole game," ask one question -- where's the state? Who's tracking whether the door is still shut? If the answer is "the neural net," it's a demo. If the answer is "a real engine underneath," it might be a product.
Vestra: That one question will save you a lot of hype for a long time.
Eris: If this made the world-model noise click into place, do us a favor -- follow or subscribe, drop a like, and share it with the one friend who keeps sending you "AI can make video games now" clips.
Vestra: And leave us a comment with this: name a game you'd want to repaint live, and the style you'd type in. Cyberpunk Mario Kart, a watercolor open-world -- we want to hear it.
Eris: And every story we touched today is written up over at Ground Truth, groundtruth.day -- the distillation fight, the math counterexample, all of it, one page each, every day.
Vestra: We read the papers so you get the commute back. See you tomorrow.