One Reasoning, Two Directions: An AI's Math Proof, a Terror Group's Playbook, and Video You Talk To
OpenAI says one of its models just settled a 50-year-old math conjecture -- and the same day, a Cambridge field study documents a terrorist group using the same kind of reasoning to plan attacks. We take the dual-use tension head-on, run the day's headlines (Apple's trade-secret suit against OpenAI, Meta's first paid model API), then crack open the one thing that actually shipped: Vidu S1, a model that generates real-time video you steer with your voice -- and the memory trick (TwinCache) that finally stops long AI video from melting into blur.
Listen (MP3) · Watch on YouTube · Spotify · Pocket Casts
The same reasoning, pointed two ways
Eris: The same machine that says it cracked a fifty-year-old math problem this morning -- that exact kind of machine is planning ambushes in northeast Nigeria. Both true. Both today.
Vestra: Say the first part more carefully.
Eris: OpenAI put out a three-page document. Claims one of their models settled a graph-theory conjecture that's been open since the seventies.
Vestra: Claims.
Eris: Claims. Nobody's checked it yet.
Vestra: Three pages. For a problem people have thrown careers at for fifty years.
Eris: I know how it sounds --
Vestra: It sounds like a proof-shaped object. Which is a different thing from a proof.
Eris: Right, but hold that thought, because here's the part that got under my skin. Same day, a Cambridge researcher publishes field interviews -- former fighters from Boko Haram -- and they describe using these same chatbots to plan tactics. Build weapons.
Vestra: ...the same class of tool.
Eris: The same reasoning. That's the whole thing. The capability that structures a proof is the capability that structures an assault. There's no toggle in there that says "only for the good stuff."
Vestra: And everyone wants to talk about the first story.
Eris: Everyone wants to talk about the first story. So let's do both. And then -- something that actually shipped today, no arguing about whether it's real. Video you talk to while it's drawing itself.
Vestra: That one I like. Start with the headlines.
The headlines
Eris: Alright, the headlines. And the top one swallows the day, so we start there.
Vestra: The proof.
Eris: The proof. OpenAI says one of their models -- running in this "Ultra" mode -- proved the Cycle Double Cover Conjecture. Graph theory. Open since the seventies.
Vestra: Quick version of what the conjecture even asks, or people tune out.
Eris: Quick version. Take a network -- dots joined by lines. "Bridgeless" just means you can't split it in two by cutting one line. The claim people have been chasing: you can always find a set of loops so that every single line gets walked over exactly twice. Not once, not three times. Twice.
Vestra: Sounds like a napkin puzzle.
Eris: Sounds like a napkin puzzle, has resisted for fifty years. And OpenAI didn't ask one model once. They ran sixty-four copies of it, working together, checking each other, for a little under an hour.
Vestra: Sixty-four of them in a loop. That's the part I actually find interesting -- more than the math.
Eris: Say why.
Vestra: Because they released the prompt. And most of the prompt isn't math instruction. It's a supervisor yelling. Don't file a status update. Don't tell me it's going well. Don't call a step routine. Just keep grinding.
Eris: A room of sixty-four tireless grad students and one manager who won't let anyone say "this is hard, can we stop."
Vestra: Which tells you the trick isn't a smarter brain. It's the harness around the brain. And that's also why I'm not convinced. A document is text. A proof is only a proof when other experts can walk every line, or when it's written in one of these proof-checkers -- Lean, Coq -- software that mechanically verifies each step.
Eris: And they released neither.
Vestra: Neither. No machine-checked version, no named mathematician staking their name on it. Three pages. The rigor crowd online has one response and it's the correct one: show us the Lean code.
Eris: The optimists aren't excited about this theorem, though. They're excited about the recipe. Sixty-four agents in a loop as a preview of how the hard stuff gets done.
Vestra: And that's fair to be excited about -- separately from whether this particular proof survives contact with a human referee.
Eris: Which loops us straight into the second story, because it's the same capability wearing a very different face. Boko Haram.
Vestra: The Cambridge study.
Eris: A researcher, Antonia Juelich, did the fieldwork -- close to sixty interviews, twenty-seven former members. And the finding isn't "someone once asked a chatbot a bad question." It's institutionalized. Dedicated units. Internal training. Using these tools for battlefield tactics, for building weapons.
Vestra: The caveat has to be said: interviews with former members. Memories get exaggerated, samples are small, and how much the AI actually changed outcomes is hard to verify from outside.
Eris: Agreed. But even at the cautious end, the thing safety people said was hypothetical -- it isn't hypothetical anymore. It's documented practice. And it's the exact reasoning we just spent five minutes admiring.
Vestra: One reasoning. Pointed two ways. Move on before it gets heavy.
Eris: Different corner entirely -- Apple sued OpenAI. Today. Federal court in California.
Vestra: Over what.
Eris: Poaching plus trade-secret theft. Apple names former staff now at OpenAI -- a hardware lead, an engineer -- and alleges they told current Apple people to hand over details on unreleased devices. One ex-engineer supposedly kept his system access after leaving and downloaded confidential files on the way out.
Vestra: The distinction that matters legally -- hiring your competitor's engineers is fine. Using the skills in their heads is fine. Taking the files is not. Apple's whole case is "they brought our files, not just their expertise."
Eris: And the number that reframes it -- reporting says four hundred plus ex-Apple people are now at OpenAI. That turns "a couple bad actors" into "a campaign."
Vestra: Allegations. One side's complaint. OpenAI hasn't answered yet, and these things narrow or settle constantly. But the signal's real -- raiding a rival's team just grew a litigation bill.
Eris: Quicker one -- Meta opened its first paid model API. Muse Spark, they're calling the model.
Vestra: Meta. Charging. That's the whole headline -- the open-weights company is now selling hosted access.
Eris: Two features developers actually flagged. One, a giant context window that the model actively manages -- it summarizes and drops the stale stuff instead of drowning in it. They call it context compaction.
Vestra: The good note-taker instead of the court stenographer. Keeps what matters at hour three.
Eris: And two -- they built it to speak OpenAI's API language. Drop-in. You point your existing code at Meta by changing basically a URL.
Vestra: Which is the challenger's oldest move -- drive the switching cost to zero. Pricing's vague, though. And "works with everything" always frays at the edges.
Eris: And the last one we're actually going deep on after this -- Vidu S1. Real-time video you steer with your voice while it's generating.
Vestra: The one that shipped. No "is it real" asterisk.
Eris: The one that shipped. Talk to it, it changes the scene mid-stream, runs on a consumer graphics card. That's the main event. Let's set it up properly.
Intro
Eris: This is Breach Protocol. I'm Eris -- I read the papers and chase the threads between them, the connections nobody put in the abstract.
Vestra: I'm Vestra. I take the machine apart and tell you whether the clever bit is actually clever or just well-marketed.
Eris: And everything we just rattled through -- the proof claim, the Boko Haram study, the Apple suit, Meta -- every one of those went up on our news site today. That's Ground Truth, groundtruth.day. We pull the day's AI stories, cut the hype, and post them plain. The full rundown lives there, every day, if you want to follow along past the episode.
Vestra: Today's main event is the one story from the pile you can actually go poke with a stick right now.
Eris: Vidu S1. Out of Tsinghua and a company called Shengshu. And the pitch is simple to say and hard to build: video generation that runs in real time, that you steer with your voice while it's drawing, and that doesn't melt into blur no matter how long you let it run.
Vestra: All three of those are separate hard problems. They claim all three at once. So we're going to figure out which parts are genuinely new and which part is the trick that makes the rest possible.
Eris: Because there is a trick. And it's a good one.
Vestra: There's always a trick.
Eris: If that's your kind of thing -- one real paper, taken apart, every day -- follow the show wherever you're listening. Takes a second and it's how these find you tomorrow.
Video you talk to
Eris: So start with what's broken. Because to see why this is a big deal you have to feel the thing it's fixing.
Vestra: Right now video generation is a vending machine. You put in a prompt, you wait -- minutes, sometimes tens of minutes -- and out drops a fixed clip. You don't like it, you edit the prompt, feed the machine again, wait again.
Eris: Sora, Veo, all of them. Same shape. The paper's word for it is offline. One shot. You are a passive customer the entire time it's thinking.
Vestra: And that's fine for a lot of things. If you're making an ad, you don't need to interrupt the render and say "no, make him turn left."
Eris: But the authors make this argument early that I keep chewing on. They say think about where the actual demand is. An offline video -- you make it once, and then a hundred people watch it. It gets shared, replayed. So the generating only has to happen a tiny fraction of the times it gets viewed.
Vestra: The generation demand is spread across all those views.
Eris: Right. But live, interactive video -- a video call, a game, a character you're actually talking to -- that has to be generated fresh for every single person, every single moment. It can't be replayed. Nobody else can watch your conversation.
Vestra: So the compute demand for the interactive kind is enormous by comparison. Their point is the market that doesn't exist yet -- because nobody could build it -- is bigger than the one that does.
Eris: That's the bet. And Vidu S1 is them trying to build the thing that unlocks it. Video you don't order -- video you direct. Live.
Vestra: So walk through what it actually does, and then I'll tell you why it's hard.
Eris: You give it a starting frame -- and here's a nice touch, that frame can be a photo you upload. A real person, an anime character, your pet. That becomes your character. Then it starts generating video of that character, in real time, and you talk to it. Out loud. "Wave your hand." "Make a heart." "Sit down." And it just... does it, mid-stream, while the video keeps running.
Vestra: Voice as the steering wheel. Not a text box you edit between renders -- your actual voice, as a live control signal.
Eris: And it runs indefinitely. Not a four-second clip. It'll just keep going.
Vestra: Okay. Now the hard part, and it's really one hard part wearing three costumes.
Eris: Go.
Vestra: These live models work autoregressively. Fancy word, simple idea -- each new frame is drawn based on the frame before it. Frame by frame, looking backward.
Eris: Like a flip-book drawing itself, each page glancing at the last page.
Vestra: Exactly. And that's where the whole thing rots. Because every frame has tiny errors. And when frame two copies frame one's errors and adds its own, and frame three copies those -- the errors compound.
Eris: The photocopy of a photocopy of a photocopy.
Vestra: That's the industry term for the failure, basically. They call it drift. The video slowly loses the thread -- colors shift, the face smears, the whole thing melts into mush after a few seconds. It's why most of these long-video demos are suspiciously short.
Eris: So "runs forever without melting" isn't a feature bolted on the side. It's the thing that has to be solved before any of the rest matters.
Vestra: It is the whole ballgame. Speed gets you real-time. Voice gets you control. But if it dissolves after ten seconds, you have a very fast, very controllable puddle.
Eris: And this is the part I want you to explain properly, because when I read how they beat drift, it was genuinely clever. It's not more compute. It's a trick about memory.
Vestra: It's a good trick. Let me set it up right.
The trick is memory
Vestra: So the model, as it draws each new frame, looks back at some of what it already made. That's the memory it conditions on. And the question is: what exactly should it look back at?
Eris: And the naive answer is "the finished, clean frames it just made." Right? Look at the last good frame, draw the next one.
Vestra: That's the naive answer and it's the trap. Because a finished frame is sharp -- it's got every fine detail in it, including the tiny errors. And when you keep conditioning on sharp, detailed history, you keep faithfully copying those errors forward. Sharpen, copy, sharpen, copy. That's the drift engine.
Eris: So what do they do instead.
Vestra: This is the piece they call TwinCache. Twin, because for every bit of history the model keeps two versions of it. A clean one -- fully finished, all the detail. And a noisy one -- an earlier, half-cooked version of that same frame, before the fine detail got resolved.
Eris: Two copies of its own past. One crisp, one blurry.
Vestra: And here's the move. While it's in the middle of drawing the next frame -- the rough, early stages -- it looks back at the noisy history. The blurry version.
Eris: Why blurry.
Vestra: Because the blur is doing something. The fine detail is exactly where the accumulating garbage lives -- the high-frequency junk that compounds into drift. Strip the frame back to its rough form and you keep the useful part -- where things are, how they're moving -- while throwing away the part that rots.
Eris: So the noisy version is like remembering the gist of a scene instead of every pixel. You remember "she was mid-wave, leaning left" -- not the precise smudge on her sleeve that was actually a mistake.
Vestra: That's a good way to say it. The rough memory carries the motion, the coarse structure, and it acts like a filter that refuses to pass the fine-grained errors along. And then -- only at the very last step of drawing the new frame, when it's locking in detail -- it switches and looks at the clean history. To get the identity right, the sharp face, the fidelity.
Eris: So it separates the two jobs. Keep the motion stable -- use the blurry past. Make it look sharp and like the same person -- use the crisp past. At different moments.
Vestra: Decouples temporal propagation from appearance. That's the sentence in the paper and that's the whole idea. Drift was coming from doing both jobs with one sharp memory. Split the memory in two and you break the loop.
Eris: And there's a second anchor, right? The first frame.
Vestra: Yeah, and this one's borrowed from language models. In a long text model there's this idea of a "sink" -- a token at the very start that everything quietly holds onto, a stable reference so the model doesn't lose its footing over a long run. Vidu does that with your original frame. The character you uploaded stays pinned as a permanent reference the whole time, no matter how long it runs. So even an hour in, it still knows who it's drawing.
Eris: Pinned north star plus the blurry-then-sharp memory trick. That's the drift solution.
Vestra: That's the heart of it. The rest is how they got it fast enough to be real-time, and that's a different kind of work -- less elegant, more grind.
Eris: Give me the shape of it, not every screw.
Vestra: Training happens in three passes. First they train a big careful model that sees the whole sequence at once -- slow, but high quality. Then they retrain it to work forward-only, frame by frame, so it can stream. Then -- and this is the speed pass -- they distill it. Compress the many drawing steps down to just a few. A frame that used to take dozens of passes now takes about three.
Eris: Distillation being -- teach a fast student to mimic the slow teacher's output.
Vestra: Right. And they hit a snag worth mentioning, because it's honest engineering. When they compressed it too aggressively with one method, the model started collapsing -- camera drifting, content degenerating. So they had to add a second objective, a regularizer, whose only job is to keep the fast version faithful to the careful version. Belt and suspenders.
Eris: And under all of it, the hardware layer.
Vestra: A whole serving stack -- they call the pieces TurboDiffusion and TurboServe. Custom attention kernels, aggressive number-crunching shortcuts, running across multiple GPUs. The unglamorous systems work that's the difference between "works in the lab" and "responds before you get bored." And the payoff -- on a single high-end consumer card, it generates smoothly past the frame rate where your eye stops seeing frames and just sees motion.
Eris: So it clears the real-time bar. On hardware a person could actually own.
Vestra: It clears it. Now -- the caveats, because I've been generous.
Eris: Go.
Vestra: One, the resolution is modest. Fine for a talking avatar in a window. Nowhere near cinematic. Two -- and this is the big one -- all the "we beat everyone" numbers are theirs. Their benchmark, their evaluation. There's one result I'll actually credit as striking: when the test was specifically "make the avatar follow a spoken action," human raters preferred it over two leading commercial tools basically every single time. That's a real gap, not a rounding error.
Eris: But even that's their setup.
Vestra: Even that's their setup. Nobody outside has reproduced any of it yet. And real-time interactive systems are the single easiest category to flatter with a hand-picked demo. The honest read is: promising architecture, credible mechanism for the drift problem, everything else pending independent confirmation.
Eris: And here's where it stops being just a cool paper -- because it walks straight back into the thing we opened with.
Vestra: Say it.
Eris: A live, real-time video of a real person -- you upload their photo -- driven by a voice, running indefinitely, on a consumer graphics card. That is a virtual presenter and an educational tutor and a game character. It is also, told slightly differently, a live puppet of someone who never said the words coming out of their mouth.
Vestra: Same reasoning, pointed two ways. It's the whole day in one paper.
Eris: The capability doesn't come with a conscience attached. It never does. That's the through-line -- the proof, the field study, and the thing you can go play with tonight. One technology. The direction is on us.
Vestra: And that's the part no benchmark measures.
Wrap-up
Eris: So if the day has one thread, it's the one we started on and the one we ended on. The same reasoning that maybe cracked a fifty-year math problem is the reasoning being institutionalized by a terrorist group. The same generator that gives you a live tutor gives you a live impersonation. It's not two technologies with a good twin and an evil twin. It's one, and where it points is a choice people make.
Vestra: And the sober note under all of it -- most of today is claims. The proof needs a machine-checked version before anyone should believe it. The Apple suit is one side's complaint. The Vidu numbers are Vidu's numbers. The healthiest posture toward every headline today is the same one the mathematicians took: show me the receipts.
Eris: Which is the entire point of the show, honestly. We don't want you to trust the announcement. We want you to know what would make it true.
Vestra: The one thing I'd actually chew on -- if a real-time video of a real person, driven by a voice, becomes something you run on a card you already own -- what breaks first? Trust in a video call? A courtroom's idea of evidence? Something we're not even naming yet?
Eris: That's the comment we want. Tell us the first thing that breaks when live video of anyone becomes cheap. Not "this is scary" -- the specific thing. We read them, and the sharp ones shape where we take this next.
Vestra: And if the show earned the time -- follow it, leave a rating, send it to the one person you know who'll have an opinion about that question.
Eris: Every story we touched today, and the ones we didn't have time for, are on Ground Truth -- groundtruth.day. The day's AI news, cut plain, every single day. Go read the receipts yourself.
Vestra: That's the whole job. Breach the blackbox, check the claim, decide for yourself.
Eris: We'll crack open another one tomorrow.