The Video AI That Never Watched: A Benchmark Audit vs. a New Recipe for Machine Sight
Two AI papers dropped the same day, and they read like a prosecution and a defense. First, an audit finds that about half of the field's video-understanding tests can be aced with the video deleted -- models are guessing from words, and once the shortcuts are stripped out, the best systems barely beat a coin flip. Then the counter: take a model trained to generate video, repurpose it as a perceiver, and it matches specialist vision systems using as little as a fraction of the data -- learning to see a real cat after training only on synthetic humans. We break down the reckoning underneath both, plus the day's headlines: Zig's creator torching the 'AI rewrote our codebase' story as unreviewed slop, Samsung Health's consent squeeze, LAPD dropping license-plate surveillance, and Richard Sutton's bet against frozen models.
Listen (MP3) · Watch on YouTube · Spotify · Pocket Casts
The Video AI That Never Watched the Video
Eris: Take one of these AI systems everyone says "understands video." Rip the video out. Hand it just the words of the question, nothing else -- no frames, no motion.
Vestra: And?
Eris: It scores about the same. Blind. It never needed to watch.
Vestra: On one weird question, sure --
Eris: On half of them. Across fourteen of the field's own benchmarks. Half the test, you can pass without ever seeing a single frame.
Vestra: Okay. That's -- hold on. That's not a bug in one dataset. That's the field grading itself on an exam it can answer with its eyes closed.
Eris: Right. And here's the part that got me. Strip those freebies out -- keep only the questions that actually need you to watch -- and the state of the art?
Vestra: Falls.
Eris: Falls to barely better than guessing. The best public models. Coin-flip territory.
Vestra: So all that "video understanding" progress on the leaderboards --
Eris: Might be a language model doing what language models do. Filling in the likely answer. There's a video bolted on the side that it barely looks at.
Vestra: A video-shaped hole.
Eris: A video-shaped hole. And then -- same day, second paper -- somebody says: fine, you want a model that actually sees? Stop teaching it to compare pictures. Teach it to make them.
Vestra: Make them.
Eris: Learn to generate the world. Then you'll learn to see it. And they've got the numbers to make you take it seriously.
The Headlines -- A Day of Reckoning
Eris: Alright, the headlines. And the mood today is one word: reckoning. Everywhere you look, somebody's asking -- was that thing actually checked, or did it just look right?
Vestra: Start with the one that lit up every developer forum. The Zig story.
Eris: So there's been this whole triumphant narrative -- AI rewrote a whole production codebase into a faster language, shipped it in a fraction of the time, proof the machines have taken over software.
Vestra: And Zig's own creator stood up and said: that's marketing. That's not a win.
Eris: Andrew Kelley, plus a widely-shared essay -- they name the failure precisely. It wasn't a language triumph. It was unreviewed slop. A mountain of AI code that reads perfectly, so nobody read it closely, and the subtle bugs sailed through.
Vestra: And that's the counterintuitive bit. We assume AI code is easier to review -- it's tidy, it's commented, it's confident. That polish is exactly the trap. Fluency does the work that scrutiny is supposed to do.
Eris: The cost of writing code fell off a cliff. The cost of trusting it didn't move. That's the whole story of engineering right now.
Vestra: Strongest pushback, to be fair -- velocity is real. If it ships far faster and you catch the bugs with better tooling, maybe slop is just the friction of a new normal, not a verdict.
Eris: Maybe. But the bottleneck moved from writing to trusting, and nobody's pretending otherwise anymore. Which -- honestly -- rhymes with our two big papers today, on video AI. Same disease. Output that looks right, nobody checking if it is.
Vestra: We go deep on those after. Keep moving. Privacy had a loud day.
Eris: Two of them. Samsung Health started asking people to let their most intimate data -- sleep, medication, cycle tracking -- train AI. And if you say no?
Vestra: You reportedly lose cloud sync. Which is the definition of coerced consent. "Freely given" stops meaning anything when saying no costs you a feature you already depend on. That's the kind of thing privacy law was written to catch.
Eris: And Meta did the same dance and lost it in two days. They launched Muse Image -- genuinely interesting model, we'll come back to why -- with a default-on feature quietly pulling public Instagram photos into training.
Vestra: The actors' union and a major talent agency came down on them, and within about forty-eight hours the data grab was gone. The model stayed. The cameras came down.
Eris: Forty-eight hours isn't deliberation. That's a company that misjudged the room and scrambled. And the pattern's the point -- default-on data harvesting now triggers organized pushback in days, not years.
Vestra: The technical footnote worth keeping: Muse Image is agentic. It doesn't paint in one pass -- it searches, writes code, refines its own output before it hands you the picture. Test-time compute, for images. That quietly shipped this week and got buried under the outrage.
Eris: Then the surveillance story, which cuts the other way -- a rollback that actually happened. LAPD let its automated license-plate-reader contract lapse.
Vestra: Flock. Cameras that read every passing plate and check it against watchlists. The problem is scale -- even a tiny error rate, times millions of plates, is a lot of innocent people getting pulled over. Some at gunpoint, over a misread plate.
Eris: And a major department walked away, naming false positives and civil liberties -- not budget. Surveillance almost never gets rolled back. The ratchet goes one way. This one didn't.
Vestra: Caveat worth flagging: watch whether they just re-sign with a different vendor. A lapsed contract can be procurement dressed as principle. If they replace it, nothing changed.
Eris: Fair. Couple of quicker ones. Apple shipped an on-device speech recognizer that roughly quartered its old error rate and beat the small, efficient tier of Whisper using about a third of the compute.
Vestra: English only, single-vendor benchmark -- so, grain of salt. But the story isn't accuracy, it's efficiency. Good enough recognition, running on your phone, cheap enough to just leave on. Private, instant, no round trip to a server.
Eris: And Richard Sutton -- the reinforcement learning godfather, the "bitter lesson" guy -- launched a new lab with a north star that's basically a shot at the whole industry.
Vestra: A trillion-parameter agent that learns and plans in real time on about twenty watts. The power budget of a human brain.
Eris: The bet underneath it: stop freezing your models. Today's big models train once and then they're frozen -- they don't learn from you, they run a fixed function. Sutton's whole career says the real path is agents that keep learning from their own experience, in the loop, forever.
Vestra: It's a mission statement, not a result. There's no twenty-watt continual learner today, and continual learning runs straight into the hardest open problem there is -- learn something new, tend to erase something old. But when the guy who wrote the bitter lesson bets a lab against frozen models, the field listens.
Eris: Two more and they're both about the gap between looking right and being right. OpenAI's new model, "Sol," is somehow too strict and too leaky at the same time.
Vestra: Which sounds contradictory until you see the mechanism. There's a smaller guardian model watching the big one. It bans people for benign stuff -- hardening their own website -- because defensive and offensive security look identical on the surface, and a small pattern-matcher can't read intent.
Eris: And it's so aggressive it flags a security guide for being too good. Too professional to be innocent.
Vestra: Meanwhile the UK's safety institute found it's still jailbreakable -- wrap the bad request in the right story and it walks right past. The core model outran the thing that's supposed to police it. That's the whole alignment problem, live, in a shipping product.
Eris: And last -- two threads, same shape. Two hundred-plus experts, Nobel laureates among them, signed a statement urging governments to prep for AI job displacement now, before it hits, not after. And separately, physicists and mathematicians are hitting a "verification lag" --
Vestra: -- AI proposing answers to hard problems faster than humans can check them. The credible one: a theoretical physicist said Anthropic's model handed him a mathematical bridge that unblocked six months of stuck research. The shaky one: a fifty-year-old math problem supposedly cracked, circulating with no peer-reviewed proof.
Eris: And that's the through-line for the entire day. Generation got fast. Verification didn't. Whether it's code, video, or math proofs -- the machine can produce faster than we can check. Every story today is some version of that.
Vestra: Which is the perfect setup for the two papers. Because one of them measures exactly how badly we've been fooling ourselves about what these models see.
Who We Are, and Today's Question
Eris: So if you're new here -- I'm Eris. I read the papers, and I chase the wire between them. When two results are secretly arguing with each other, that's the thing I want to pull on.
Vestra: And I'm Vestra. I take the shiny claim apart and check whether the mechanism actually holds. I'm the one asking "but how, exactly, and what would make this false."
Eris: Between us, that's the show -- Breach Protocol. We crack open the week's AI research into something you can actually follow on your commute. No wall of jargon.
Vestra: And every story we touch here goes up daily on our news site, Ground Truth -- groundtruth.day. The full rundown, every headline from today, in plain language. That's the place to follow the whole thread.
Eris: Today's main event is a rare gift: two papers, same day, that read like a prosecution and a defense. The first one goes: most of what we call "video understanding" in AI is a magic trick -- the model isn't watching, it's guessing from the words.
Vestra: And the second one goes: okay, you want a model that genuinely sees? Here's how -- and it's not the recipe anyone's been using.
Eris: One diagnoses the disease. The other proposes the cure. And they were posted the same day without knowing about each other. That's the episode.
Vestra: We'll take them in order. The audit first -- because you have to see how bad the problem is before the fix lands.
Eris: If that's your kind of thing -- two ideas colliding, taken apart carefully -- follow the show wherever you're listening. It's the fastest way to keep the thread going.
The Audit -- Most Video AI Isn't Watching
Eris: Okay. The prosecution. This is a group out of a Korean university and NAVER's cloud lab, and they did something refreshingly humble. They didn't build another benchmark. They audited the ones we already have.
Vestra: Which is the right instinct. Everyone keeps shipping new video tests. Nobody stops to ask whether the old ones measure what they claim.
Eris: Fourteen of them. Perception tasks, reasoning tasks, clips from a few seconds to hours long. And they run this diagnostic suite that basically asks one question over and over: does this question actually require the video?
Vestra: And how do you test that? Because that's the clever part.
Eris: You attack it from every side. First one -- just blindfold it. Delete the video entirely. Give the model only the question and the multiple-choice options. Nothing to look at.
Vestra: And a blindfolded model should score at chance. Random guessing.
Eris: Should. It scores well above it. Comfortably. On the question text alone.
Vestra: So the question is leaking the answer. Either the options give it away, or the model knows what usually happens in this kind of clip.
Eris: Exactly that. "Is the person jumping?" -- in a category where people jump all the time, you say yes and you're right most of the time. You never watched. You played the odds.
Vestra: Right, okay. But blind is the crudest test. What about time -- the thing that supposedly makes video video?
Eris: They went after that too, and this is the one that got me. Take the frames. Shuffle them. Totally scramble the order.
Vestra: Destroy the timeline.
Eris: Destroy the timeline. If the model understood the event -- first this happened, then that -- scrambling should wreck it. It barely flinches.
Vestra: Mm. So it's not watching a sequence. It's glancing at a pile of frames and pattern-matching.
Eris: And they confirm it another way -- give it only the single middle frame. One still image. Does nearly as well. So the "temporal reasoning" was mostly spatial recognition wearing a costume.
Vestra: Let me push on the method, though, because this is where audits usually go wrong. If a model happens to guess right with the video removed, that doesn't prove the question is broken. It could just be lucky, or the model could be strong.
Eris: Good -- and they anticipated exactly that. Two guards. One: they don't trust a single model. They use several, and only call a question a "shortcut" when multiple independent models all nail it blind. Consensus.
Vestra: So it's not one flaky model. It's a reliable freebie.
Eris: Two: they went back and checked -- are these flagged questions actually easy under normal conditions, with the video? And they are. The overwhelming majority get answered fine in the standard setup. So these aren't broken junk questions --
Vestra: -- they're genuinely answerable, just answerable without looking. That's the distinction that makes it damning. And they proved it's not just "hard questions," too?
Eris: They checked that directly. Compared their flagged set against the set of questions all the models get wrong -- pure difficulty. Barely overlaps. Under half. So they're not skimming off hard problems. They're finding a specific defect: the question doesn't force you to use the video.
Vestra: And the headline number falls out of that.
Eris: About half. Roughly half of all the samples, across those fourteen benchmarks, are solvable with no video and no timeline. And there's a nasty correlation on top -- the benchmarks with the most shortcuts tend to report the highest scores.
Vestra: Of course they do. The leaderboard rewards the gameable test. The easier it is to cheat, the more impressive everyone looks.
Eris: So then they do the honest thing. They strip the shortcuts out. Keep only the questions that genuinely need you to watch and track and reason across time. And they re-run the best models on what's left.
Vestra: This is the number from the cold open.
Eris: This is it. The state of the art collapses to barely above guessing. One frontier proprietary model stays meaningfully above the pack -- but even it is landing in the middle of the range, not acing anything. The rest are near the coin flip.
Vestra: So let me say the uncomfortable version plainly. A lot of the "video-LLM" progress we've been applauding is a language model with a video encoder taped to the side, and the language half is doing almost all the answering.
Eris: A video-shaped hole. Their phrase, more or less. And the worst category -- the one everybody's weakest at -- is following a long, multi-scene story. Holding a whole timeline together. That's the real frontier and nobody's close.
Vestra: Here's what I appreciate, though -- they don't just diagnose and walk away. They use the cleaned-up set as a testbed and ask: okay, what actually helps?
Eris: And two findings there are genuinely useful. First -- temporal grounding. Meaning: before you answer, go find the specific stretch of video the question is about, instead of skimming the whole thing evenly.
Vestra: Which sounds obviously good. Did it help?
Eris: Only a little, when they bolted on a normal version. But then they ran an oracle -- they cheated, handed the model the perfect, ground-truth stretch of video every time. And on the honest, shortcut-free questions, performance jumped a lot.
Vestra: And on the shortcut questions?
Eris: Almost no change.
Vestra: Because the shortcut questions never needed the right moment in the first place. That's actually a beautiful result. It says precise grounding only matters exactly when the task really is about the video -- and the reason grounding looked useless before is we were measuring it on tests that didn't need it.
Eris: The tool wasn't weak. The exam was fake. Second finding -- when to think.
Vestra: Meaning reasoning depth.
Eris: Right. They compare a model that always reasons hard versus one that just answers, and always-thinking barely wins. But then, another oracle -- imagine a perfect switch that picks the better mode per question, think when you should, answer fast when you shouldn't.
Vestra: And that switch alone --
Eris: Nearly catches the frontier model. Just from knowing when to deliberate. Their line, and I love it: choosing when to think can matter as much as raw scale.
Vestra: Which is a quietly radical thing to say in a field that answers every problem with "make it bigger." Now -- the caveats, because we owe them. This is fresh. Not independently reproduced yet. And the exact "half" figure depends on which fourteen benchmarks they picked -- a different basket gives a different fraction.
Eris: All fair. But the failure mode itself -- models passing visual tests without looking -- that's old and well-documented. This is just the most systematic nose-rubbing the video field has gotten.
Vestra: So the disease is named. Fluent guessing dressed up as perception. Which is the exact same sentence we said about the code slop this morning. Now -- what's the cure.
The Cure -- Learn to Make It, Then You'll See It
Eris: The defense. This one's out of DeepMind, and it starts from a question that sounds almost philosophical. Language got its foundation model from one trick -- predict the next word, at massive scale, and general intelligence sort of fell out. What's the equivalent trick for vision?
Vestra: And the field's had a few answers. The CLIP answer -- contrastive learning, teach a model to match pictures to captions, pull the right pairs together, push the wrong ones apart. The masked answer -- hide part of an image, predict what was there.
Eris: And this paper says: both of those are the wrong teacher. The right one is generation. Learning to make video.
Vestra: Okay, unpack why they think that's not just a slogan.
Eris: Their argument is that to generate convincing video -- frame after frame, physically plausible -- a model is forced to internalize the actual structure of the world. How objects move. That things don't teleport. How light lands on a surface. How a scene stays coherent in three dimensions as time runs.
Vestra: Whereas "match this picture to this caption" only teaches you a shallow association. You can learn that a beach photo goes with the word beach without understanding a single thing about depth or motion.
Eris: Right. Generation forces the deep model. So their move -- and this is the elegant part -- they take a big pretrained text-to-video generator and they don't use it to generate anything.
Vestra: They repurpose it. Into a perceiver.
Eris: Into a feed-forward perceiver. Normally these video models denoise -- they run dozens of slow steps, starting from noise, gradually sculpting a clip. That's expensive and it's a generator. They shortcut all of it. Feed in a clean video, fix the process to a single pass, one shot, and read the features straight out.
Vestra: So they've turned a slow, iterative painter into a fast, single-glance feature extractor. And there's a specific trick to make the output line up -- they flip the sign of what the model predicts, because of how it was trained to move from noise toward image. Little detail, but it's the thing that makes the reuse actually work.
Eris: And then one model, one set of weights, does everything -- depth, surface normals, camera pose, cutting out objects, following a spoken description to segment the right thing, 3D body pose. You just change the text prompt to pick the task.
Vestra: This is the claim I want to stress-test. "Matches the specialists." Because specialists are hard to beat -- that's the whole point of specializing.
Eris: And it does. Across that spread of tasks it lands at or above dedicated models that each do only one thing. But that's not the headline for me. The headline is what it cost them to get there.
Vestra: Data.
Eris: Data. It reaches the same league as the leading specialist systems using somewhere between seven times and several hundred times less training data. Depending on the task -- but at the extreme, hundreds of times less.
Vestra: Now that -- if it holds -- is the whole ballgame. Because data is the tax on everything. Hundreds of times less isn't a tuning win, it's a different economy. But I want to know it's the generation part doing the work, and not something boring, like they just had a bigger model or a better dataset.
Eris: They ran exactly that control. Same finetuning, same data -- swap the generative backbone for the other big self-supervised approaches, the masked-prediction one, the joint-embedding one. The generator wins clearly.
Vestra: So it's the objective, not the scale.
Eris: And to nail it -- they tried starting the same architecture from scratch, no pretraining. The learning curve is basically flat. It goes nowhere. The more of the pretrained generative layers they keep, the better it gets. The generative knowledge is load-bearing. Take it away and there's nothing there.
Vestra: Okay. That's a real result. What's the emergent behavior everyone's quoting?
Eris: This is the part that made me sit up. They trained it -- for most tasks -- purely on synthetic video. Computer-graphics humans. Rigged 3D characters, motion-capture animations, rendered in Blender. Fake people.
Vestra: Trained on fake humans.
Eris: Trained on fake humans, single figure per clip. And it generalizes to real footage. To scenes with several people at once. And -- this is the kicker -- to things it never saw at all. Animals. Robots. It'll pull the depth and the pose off a cat.
Vestra: A cat. Trained only on synthetic humans.
Eris: They point out the whiskers come out sharper than anything in the training data. It's not parroting what it saw. If fake humans teach you to see a real cat, you didn't memorize appearances -- you learned something about how bodies and surfaces and geometry work in general.
Vestra: Which is the strongest evidence for their thesis, honestly. The generative pretraining gave it a little world model, and perception is just reading off of it. That's a clean story. So let me do my job and find the seams.
Eris: Please.
Vestra: One -- they're candid about a failure, which I respect. When they cram all the tasks into one joint model, some of it regresses. The 3D pose in particular degrades. Their own read is that the bolted-on machinery for the point-coordinate tasks disrupts the native attention the generator came with. The lesson they draw -- touch the pretrained backbone as little as possible -- is real, but it also means the "one model does everything, cleanly" story has rough edges.
Eris: Fair. It's not free.
Vestra: Two, and this is the one I actually want on the record. This is fresh arXiv, not reproduced, and a claim of hundreds-of-times data efficiency is exactly the kind of extraordinary result that sometimes shrinks under independent replication. Hold it loosely until someone else runs it.
Eris: Agreed. Log the candidate, wait for the check -- which is, funnily enough, the exact discipline the verification-lag story asked for this morning.
Vestra: And three -- the one nobody's saying. Everybody's framing these two papers as disease and cure. Video-Oasis: models can't really see. GenCeption: here's how to make them see. But they're not actually operating on the same layer.
Eris: Say more.
Vestra: GenCeption is low-level perception. Depth, normals, where the body is, cut out this object. Video-Oasis is testing high-level reasoning -- follow a multi-scene plot, infer cause and effect, hold a narrative across an hour. Learning to see a cat's geometry beautifully does not, by itself, mean you can reason about why a character did something forty minutes into a film.
Eris: So the cure treats the foundation, not the whole illness.
Vestra: Right. It's necessary, plausibly. Maybe you can't reason about a scene you can't even perceive. But it's a floor, not the building. Nobody's shown that better perception automatically climbs into better narrative understanding.
Eris: And yet -- here's why I still think they belong in the same breath. Video-Oasis basically says: the reason models fake it is they never built real spatial and temporal grounding, so they lean on text priors instead. GenCeption says: here's a way to actually build that grounding, cheaply, by learning to generate the world. One shows the hole. The other shows a way to start filling the bottom of it.
Vestra: I'll take that framing. Foundation, not cure. And both of them, in their own way, are saying the same thing the whole day has been screaming.
Eris: Which is?
Vestra: Stop trusting the surface. A model that sounds fluent, a benchmark that reports high, code that reads clean -- none of that is evidence it's real. Somebody has to check whether the thing actually did the thing. That's the whole reckoning.
The Reckoning -- Stop Trusting the Surface
Eris: So here's the day, tied off. One thread runs through every single story. Generation got cheap. Verification didn't.
Vestra: The Zig fight -- code that reads clean but nobody checked. The video audit -- benchmarks that report high but the model never watched. The math proofs -- answers arriving faster than anyone can referee them. Same shape, three fields.
Eris: And the one hopeful note in the pile was GenCeption -- not because it's a miracle, we spent ten minutes poking holes in it -- but because it points at the actual fix. If you want a model that really understands the world instead of faking it, make it learn the world deeply. Learn to build it. Then it can see it.
Vestra: And if it's real, it does that for a fraction of the data. Which we'll believe when someone reproduces it. Log the candidate. Wait for the check.
Eris: That's the posture. For code, for video, for proofs, for us. The fluency is not the evidence.
Vestra: We want something from you, though -- genuinely. We keep coming back to this idea that AI shifted the hard problem from making to trusting. We want to know where you're feeling it. In your actual work.
Eris: Yeah -- drop us a comment: what's something an AI handed you lately that looked completely right, and turned out to be quietly wrong? Code, a summary, a fact, a citation. The thing that fooled you for a second. We read those, and the good ones shape future episodes.
Vestra: If this was worth your commute -- follow the show, wherever you're listening. Give it a like, it genuinely helps people find us. And send it to the one person you know who keeps forwarding you AI hype -- they need it most.
Eris: And every story we touched today, in full, plain language, one page -- it's up on our news site, Ground Truth. That's groundtruth.day. New stories every day, same voice, same standard: what was actually checked.
Vestra: Because that's the whole name of the show. Ground truth. The thing that's real after you stop trusting the surface.
Eris: We'll see you tomorrow.