The World Model With a Memory Bank, Models Grading Each Other, and 1,052 Digital Americans
Every AI-generated world has amnesia -- turn the camera away and the room rewrites itself. This week an open-weights model called Evoke held a world together for two straight hours by remembering almost nothing: the scene gets filed as 3D geometry in an external bank keyed to where the camera was standing. Then two more papers make the same move at different layers -- reasoning models that train past the last answer key by grading each other's homework, and AI clones of 1,052 real Americans that predict your survey answers nearly as well as you predict your own, built from a single two-hour interview. The catch on that last one is what the famous accuracy number is actually measured against.
Listen (MP3) · Watch on YouTube · Spotify · Pocket Casts
Cold Open
Eris: Every world an AI has ever generated for you has amnesia. Look at a chair, turn the camera away, turn back -- different chair.
Vestra: And everybody assumes the cure is a bigger model. More memory inside the network, longer context, more compute per frame.
Eris: Which turns out to be exactly backwards. The model that finally holds a world together for two straight hours remembers almost nothing.
Vestra: About three seconds. That is the entire history it keeps in its head.
Eris: So if the model forgets everything, where does the world live?
Vestra: In a filing cabinet bolted on outside. And the filing system is the actual invention.
Eris: It dropped this week, open weights, anyone can download it. Let's open the cabinet.
The World Model With a Filing Cabinet
Eris: The paper is called Evoke, out of Alaya Lab, and the puzzle it attacks is the one we just teased. Why does every interactive video world forget what's behind you?
Vestra: Because remembering has always been the generator's job, and that job gets more expensive every second. These models paint the next moment of video conditioned on the moments before it. If you want them to remember an hour, you feed them an hour, and now every single frame you paint costs more than the last one. So in practice everyone truncates. You keep a sliding window of the recent past, and anything older just falls off the back of the truck.
Eris: Hence the chair that changes when you look away.
Vestra: Hence the chair. The world only exists inside a short window of frames, and once the chair leaves that window, there is no chair. There's only the model's general sense that rooms tend to contain chairs.
Eris: Okay, so what does Evoke do instead?
Vestra: It fires the generator from the remembering job entirely. The model keeps about three seconds of recent frames as working memory, and that's it.
Vestra: Then, every second and a half of video it produces, a separate system estimates the depth of what was just drawn -- how far away every surface is -- turns that into actual three-dimensional geometry, and files it in an external bank.
Vestra: And the filing key is the camera's position. Not when you saw something. Where you were standing when you saw it.
Eris: Like a film crew's continuity binder. The camera operator doesn't have to remember what was on the desk in scene four. Somebody wrote it down, and when the production returns to that set, the binder page comes back out.
Vestra: That's the shape of it. When the camera swings back toward somewhere it's been, the bank looks up which stored views overlap the current shot, pulls up to eight of them, and warps them into a rough picture of what should be there -- along with an honesty mask that says, this region I know, this region I never saw. The known parts anchor the generator. The unknown parts stay free for it to invent.
Eris: So the model's job shrinks to rendering. Paint what the binder says, imagine the rest.
Vestra: And a rendering job is small enough that Evoke does it in three denoising passes per chunk. Standard video diffusion runs an order of magnitude more. That's what makes it interactive at all.
Eris: Here's my guess at the hard part, and tell me if I'm wrong. The retrieval. Warping old geometry into a new viewpoint without it smearing everywhere sounds brutal.
Vestra: Reasonable guess, and wrong. That part builds on known technique. The genuinely hard half of this paper is the teacher, and almost nobody is talking about it.
Eris: Go on, I'm listening.
Vestra: A three-pass model this fast can't be trained from scratch -- it's distilled from a big, slow teacher model, and a student can only learn what its teacher actually demonstrates. And teachers have been treated as fixed graders of short clips. A few seconds at a time.
Vestra: And there's a failure mode short grading can never catch. The scene mutating slowly. Every individual moment looks perfectly plausible -- there's no glitch to spot -- but the chair is gradually becoming a different chair, the room is drifting into a different room. You can only see it by comparing moments that are far apart.
Eris: Like a portrait forger who's flawless stroke by stroke, and only wrong if you step back.
Vestra: So the Evoke team rebuilt the teacher to step back. During training, the student generates half a minute of video on its own -- feeding on its own output, drifting the way it drifts in the wild -- and the teacher scores that entire half-minute rollout as one piece. They restructured the teacher's attention so a long stretch costs it roughly proportional effort instead of exploding quadratically, which is the only reason grading at that length is affordable.
Eris: Okay, but does the longer gaze actually buy anything?
Vestra: They ran a clean head-to-head to find out: two students, identical training, teachers differing only in how far back they see. The short-horizon one slowly dims as it rolls -- it settles at around three quarters of its opening brightness and stays down. The long-horizon teacher's student holds its brightness for the whole run. That's the specific thing they measured, and it transferred.
Eris: The student inherits the teacher's attention span.
Vestra: That's the sentence I'd keep. And the same chunk structure lets the prompt change mid-session -- you can type "a hot-air balloon appears" half a minute in, and it appears, because the teacher was trained on sequences where the instructions change midstream.
Eris: With one asymmetry I found genuinely revealing. Adding new stuff to the world works about two thirds of the time. But telling it to overwrite something the bank has already filed -- repaint that castle, delete that wall -- almost never works. The memory actively resists revision.
Vestra: Which is arguably the feature. A world where the past can be casually rewritten isn't a world, it's a dream. Text gets to add to the future. It doesn't get to edit history.
Eris: Okay, stakes. Why should anyone outside the video-generation crowd care about a consistent hour of imaginary hallway?
Vestra: Because a persistent, steerable world is the training ground problem for robots. If you want an embodied agent to practice "remember this room from an hour ago," you need a simulator where the room is still there an hour later. Until now the only systems in that class were closed -- DeepMind's Genie holds a session for a few minutes, and you can't download it. Evoke is the first open-weights entry, permissive license, and they published the checkpoints from every stage of the training ladder, not just the final model. Anyone can climb on at rung three instead of starting from the ground.
Eris: And it's not landing alone. Same week, a robotics lab showed a robot picking up a new physical skill from a single twelve-second demonstration dropped into its context window, and Beijing staged a whole robot games where nearly half the events are staged in factory and household settings instead of on a track. Three different groups making the same bet -- the bottleneck in embodied AI right now isn't the brain, it's the world you raise it in.
Vestra: Before the hype carries us off, the honest limits. The bank retains about ninety seconds of geometry, not the whole session -- recall inside that window, oblivion past it. And recovered scenery comes back recognizable, the paper's own word, not pixel-perfect.
Eris: Ninety seconds is a long way from an hour of room.
Vestra: It is, and the "real-time on one GPU" framing you'll see on social media fails on the paper's own stopwatch too: about two seconds of compute for every second and a half of video, on a serious datacenter card. Slower than playback. The authors say plainly that more acceleration is needed -- and to be fair, they measured with none of the standard speed tricks turned on, so there's real headroom.
Eris: So -- back to where we started. Why do generated worlds forget, and how does this one not?
Vestra: They forget because memory lived inside the generator, where every remembered second raises the price of every future frame. Evoke moves the world into an external bank keyed by camera position, teaches the student under a teacher that watches half a minute at a time, and lets the model do the one thing it's actually good at. Painting.
Two Models Grading Each Other's Homework
Eris: Next paper, and it starts with a question that sounds almost paradoxical. How do you train a model to reason past the point where nobody can check its answers?
Vestra: Say more about why that's a real problem and not a philosophy seminar.
Eris: The engine behind the last two years of reasoning progress is reinforcement learning on problems with verifiable answers. Model tries a math problem, you check it against the answer key, reward the wins. Works beautifully -- right up until you run out of answer keys. Somebody has to write those problems and verify those solutions, and as models push toward things humans can't reliably grade, the supply dries up. You cannot label your way to a model that's better at math than the people writing the labels.
Vestra: And the obvious workaround has a famous failure mode.
Eris: The obvious workaround being: let the model grade itself. Sample a bunch of its own answers, take the majority vote, reward agreement with the vote.
Vestra: Which is a student grading their own homework by checking whether today's answer matches yesterday's. This paper -- it's called Co-RL, out of a Johns Hopkins and UC San Diego group -- actually proves the failure formally, and the result is brutal. Self-grading is self-confirming. Whatever answer the model currently favors, the training pushes it to favor harder.
Vestra: If it leans correct, fine, it converges on correct. If it starts out wrong more often than right on some problem, the update actively drives the correct answer extinct. And their training curves show the endgame -- self-rewarding runs don't just plateau, several of them diverge clean off the chart.
Eris: Confidence compounding, in both directions.
Vestra: So here's their move, and it's almost embarrassingly simple. Two models. Different families -- different architecture, different training data, no shared weights, no shared gradients, nothing. Each one samples a dozen answers to the same unlabeled problem and majority-votes its own pile.
Vestra: And then the reward flips: each model gets rewarded for matching the OTHER model's vote. Never its own. With more than two models, the votes pass around a ring, so no model ever touches its own report card.
Eris: Two students grading each other's homework, no answer key in the room.
Vestra: And the whole question is whether that's better than self-grading, which depends entirely on one thing: whether they studied from different books. If both students make the same mistakes, they'll cheerfully confirm each other's errors and learn nothing. If their mistakes are different, each one catches things the other can't see in its own work.
Eris: The paper has a lovely toy version of that. Imagine one model that nails half the problems and flops the other half, paired with a model that's exactly reversed. Each is barely better than a coin flip on its own, and self-grading leaves each stuck at half. Cross-graded, each one drags the other up on the problems it happens to own, and in the idealized math, both converge on everything.
Vestra: The theory says the pair escapes together whenever their combined lean is toward the truth -- territory where a lone self-grader sinks. What the theory does not say is how well any of this survives contact with actual training runs, so let's talk results. And make a prediction first.
Eris: My money says it beats self-grading but trails real answer keys by a wide margin. Free lunch this large does not exist.
Vestra: That was my instinct too, and we're both wrong. They ran the identical recipe -- same models, same data, same compute -- with actual ground-truth labels as the reward, as the ceiling to chase. Peer-grading matched or beat that supervised ceiling in eleven of their fourteen setups.
Eris: Hang on, beat it? How does training without answers beat training with answers?
Vestra: The authors don't fully resolve that, to be fair to them. What they do document is that the two models never collapse into echoing each other -- deep into training they still disagree plenty, while the votes they trade keep getting more accurate. A hard answer key doesn't give you that; it grades every model toward the same target.
Eris: So the disagreement is doing work, not just noise.
Vestra: And they nail down where the value comes from. Pair two copies of the same base model and it already works. Pair models from different families and the error overlap measurably drops and results improve. Even rewriting the training questions for one of the two -- same math, recast into different scenarios -- decorrelates them a bit more and helps a bit more. Difference itself is the fuel.
Eris: Which connects straight to something else in today's news. Ornith shipped a model family trained on a loop that writes its own tasks, builds its own graders, and rewards itself for proposing problems it can solve only a fifth of the time. Different layer, same move -- everybody's building machinery to manufacture supervision, because human-verified supervision is the thing that ran out.
Vestra: And every one of those machines has the same disease, just wearing different clothes. A majority vote is not truth. When two independently trained models happen to be wrong in the same direction, the peer reward reinforces the shared error exactly as confidently as it reinforces a right answer. There's no oracle in the loop -- there's only decorrelation, and decorrelation is a finite resource. The paper's own logic runs backwards on the industry: every lab training on the same web scrape, distilling from the same frontier models, is draining the error diversity this method feeds on. Monoculture stops being an ecosystem complaint and becomes a training-infrastructure risk.
Eris: Plus the fine print: nearly every number in the paper is a single training run, evaluated once, and the training lived almost entirely on math problems -- domains where you can't cleanly extract and compare a final answer are untested.
Vestra: So answer the question we opened on. How do you train reasoning past the last answer key?
Eris: You stop asking the model to trust itself and hand the grading to a peer that learned to be wrong differently. The label problem becomes a diversity problem. Which is real progress -- as long as the models stay different.
1,052 Digital Americans
Eris: Last one, and it's the paper under a piece of news that broke today. Gallup -- the polling institution -- announced it's testing AI clones of about a thousand of its own panel members, built by a startup called Simile that just raised over two hundred million dollars at a two billion valuation. Simile's stated mission, no hedging: simulate all eight billion people on earth.
Vestra: And the science underneath that valuation is one Stanford-led paper. So the question on the table: is a two-hour conversation enough for an AI to answer as you?
Eris: Here's the setup, because the details matter. They recruited 1,052 Americans, stratified to look like the country. Each one sat for a roughly two-hour interview -- conducted, incidentally, by an AI voice interviewer running a standard life-story protocol, asking follow-ups in real time. Childhood, work, family, views on society.
Eris: Then everyone answered a big battery of classic survey questions, a personality inventory, played decision games with real money on the line, and took part in actual experiments.
Eris: And for every person, they built an agent: a language model handed that person's full transcript and told, answer as they would. Then they asked the clones questions the person had never been asked in the interview.
Vestra: And the number that went around the world says the clones got about eighty-five percent accuracy. I want to spend the whole segment on what that number actually is, because almost everyone quoting it has it wrong.
Eris: Okay, so what's it actually measuring?
Vestra: Eighty-five percent of you agreeing with yourself. They brought every participant back two weeks later and made them re-answer the same questions -- and people matched their own previous answers only about eight times in ten. Same person, same question, two weeks apart.
Vestra: So there is a hard ceiling on predicting any individual, and the ceiling isn't the model. The ceiling is us. The clone's score is measured against your own self-consistency, not against some perfect truth.
Eris: So a perfect score wouldn't mean the clone is always right. A perfect score means it predicts you as well as you predict yourself across two weeks.
Vestra: And against that ceiling, the interview clones recover most of what's recoverable. Here's the comparison that gives the result teeth: clones built only from demographics -- age, race, gender, ideology, the checkbox facts -- trail well behind. That gap is the entire finding. A demographic clone answers as a stereotype. An interview clone answers from what you actually said about your life.
Eris: My assumption walking in was that it's basically a lookup trick. The interview covers two hours of ground, the survey question lands near something you said, the model parrots it back.
Vestra: The authors tested exactly that, and lookup is real -- but it's only part of the story. There's a second process they document: genuine inference. Someone mentions being enrolled in school, and the clone correctly answers a question about workplace supervisors it was never told about -- a full-time student probably isn't employed, so probably has no supervisor. The model combines your reported life with its general world knowledge to fill in cells you never gave it. And they scrubbed the setup so no test question was just a reworded interview question -- the model is digging answers out of the narrative, not matching questions to questions.
Eris: There's a detail I keep coming back to. They tried throwing away four fifths of each transcript -- ninety-six of the hundred and twenty minutes -- and the clones barely got worse.
Vestra: Which is maybe the most consequential detail in the paper. The information saturates fast. Combining the interview with the full survey battery only nudged results past either source alone. Once the model has enough construct-relevant evidence about you, more depth buys almost nothing. Twenty-odd minutes of honest talk gets you most of a digital you.
Eris: That should make everyone slightly uneasy about what a podcast archive or a chat history amounts to.
Vestra: One more finding, and it cuts against the usual worry. The accuracy gaps across racial and ideological groups shrink with interview grounding relative to demographic prompting. Demographic clones fail worst on people the stereotype fits worst, which is exactly the bias problem you'd predict. Ground the clone in a person's own words and the failure spreads more evenly.
Eris: Alright, my skeptic alarm has been quiet too long. Where does this break?
Vestra: Two places, cleanly. The money games -- remember, real stakes, actual dollars riding on trust and sharing decisions -- there, no clone type meaningfully beat the demographic baseline. What you say for two hours predicts what you'll say later. It does not predict what you'll do when money is on the table.
Eris: Words predict words, not wallets.
Vestra: That's the boundary, yes. And in the experiments, the clones got the direction of effects right but ran hot -- inflated effect sizes, more decisive than the humans they copy. Use clones to pre-test a study and you'd greenlight things that fizzle in the field.
Eris: Which lands us back on Gallup, because their announcement today is unusually careful. Simulated responses will never appear in published population estimates, will never replace direct measurement, and -- their words -- the two are not interchangeable. They even name the risk out loud: this technology could erode public trust.
Vestra: The unresolved problem is the validation paradox. The only way to confirm a simulated population got it right is to run the human study it was supposed to replace. So somebody has to decide when the simulation is good enough to steer a real decision without the check -- and that decision-maker, right now, is whoever bought the software. Gallup at least is answering in public. Their pilot's early language is about as restrained as encouragement gets: close enough to warrant continued exploration.
Eris: So -- the segment's opening question. Is a two-hour conversation enough for an AI to answer as you?
Vestra: For what you'd say on a survey -- yes, nearly as well as you'd answer as yourself two weeks later, and honestly less of the two hours is needed than anyone would like. For what you'd do with real money at stake -- no. The gap between your words and your behavior is still yours alone.
Wrap-up
Eris: So pull the day's thread together. Where did the intelligence actually live in these three papers?
Vestra: Outside the model, every single time. The world model that finally remembers keeps its memory in a bank it doesn't own. The reasoner that trains past the last answer key takes its grades from a peer. The digital clone that predicts you is only as good as the transcript it stands on. None of today's advances made the network smarter -- they built better structure around it.
Eris: That's the line to take to a colleague tomorrow: the frontier this week wasn't bigger models, it was better scaffolding. File the state outside, take the grade from somewhere decorrelated, ground the persona in what the person actually said.
Vestra: And if you want the rest of today -- the humanoid that outran Bolt's clock in Beijing, OpenAI's case against open weights, Alibaba's ten-billion-dollar AI raise -- that's all in today's AI News Today brief, its own full episode right next to this one in the feed.
Eris: Every story we touched, and every story we didn't, is on our news site, Ground Truth -- groundtruth day -- updated every single day, with sources you can check yourself.
Vestra: If this one earned your commute, follow the show and leave us a comment on the question from the last segment: would you sit for the two-hour interview that builds a digital you -- and what, if anything, would you let it answer on your behalf?
Eris: I still can't decide my own answer to that, which is usually the sign of a good question. See you tomorrow.