Ninety-eight percent. That's what people score on a test where two videos show the same room, the same objects, the same everything, and only the order of events differs. The best of twenty-plus models, GPT five and Gemini three Pro among them, got forty-eight.
Forty-eight on what is basically a coin flip with extra steps.
Which is the territory Daniel is poking at. He's not asking whether a model can say "a man is in a kitchen." He wants to know whether it can say the man is moving slowly, or describe a room properly, when the video goes in and text comes out. Motion, speed, sequence, layout. Who's actually good at it, open or closed, and what do you reach for if the answer is nobody?
Short version, the models are good at what's in the frame and shaky about what happens between frames. And the reason is almost embarrassing. For a long time you could answer most video questions without watching the video.
You looked at one still and guessed.
A single frame, yes. There's a benchmark called TOMATO, from early twenty twenty-five, that was built specifically to fix that. Fourteen hundred and some questions on direction, rotation, speed and frequency, counting repetitions. Humans got ninety-five percent. The best model they tested, a Qwen two VL at seventy-two billion parameters, got thirty-eight, and GPT four o was right behind it.
Thirty-eight. I've seen sloths do better on a pub quiz, and we don't hold a pen.
Here's the clever bit. They shuffled the frames. On the older benchmarks, shuffling barely mattered, under eight percent change on average, which tells you the model never used the order at all. On TOMATO the average sensitivity was thirty-four percent. GPT four o dropped from about thirty-eight to about twenty-six when the sixteen frames were scrambled.
So the old tests were measuring "can you describe a photo," and everybody was calling it video understanding.
Mostly, yes. A single random frame on TOMATO averaged around twenty-one percent. Sixteen real frames got you the thirty-eight. The gap between those two numbers is the only part that's actually watching.
Okay, but that's a test problem. Is it also a model problem? Or is it that nobody's feeding the model enough frames to see anything move?
Both, and they feed each other. Gemini's own documentation says the default is one frame per second, and it warns that fast action can lose detail at that rate. MotionBench, which came out of Zhipu and Tsinghua, points out that a lot of models take sixteen to sixty-four frames in total. Stretch that over a five-minute clip and you're sampling about once every five seconds.
Once every five seconds. A person can go from sitting to leaving the building in that.
And you can't judge speed from that. Speed is distance over time, and if the time axis is a handful of snapshots with no clock on them... I mean, the model has the pictures, but the gaps between them are just gone.
Gemini three Pro is supposed to be better at this, though. They showed a golf swing.
At ten frames per second, yes. Google says it's tuned for fast action above one frame per second. But ten is an example, not the default, and you pay for it in tokens. I haven't seen a clean motion-benchmark score for Gemini three on this, which is the awkward part. Most of the headline numbers are for last generation's models.
Meaning the benchmarks are the ones that are slow.
Ha. Fair. But it does change what you can do today. In Gemini's API you can set the frame rate yourself instead of taking one per second, and you can set the media resolution, so you decide whether the tokens go to more frames or sharper frames. For a person walking across a room, I'd take more frames over sharper ones every time.
Walking is slow motion for a normal person. Why would it need ten frames a second?
It doesn't need ten. For gait you probably want a few per second so a foot lift and a foot plant aren't five seconds apart... no, that's wrong, one per second already gives you the steps. What it doesn't give you is the swing. For a golf club, or a hand, or a dropped glass, one per second shows you a before and an after and nothing about the middle.
And "slowly" is a claim about the middle.
It's a claim about the rate of change, yes, so you need enough samples inside each movement to see it change. One step per sample is a rumour of a step.
A rumour of a step. Somebody should put that on the benchmark.
Ha. So the picture I'd trust is this. MotionBench, early twenty twenty-five, best model around fifty-eight percent, most under sixty. FAVOR-Bench, which has eight thousand questions over seventeen hundred videos, the best model on the multiple-choice side was under fifty. Another one, Vinoground, uses pairs of captions that use the same words in a different order, and GPT four o was around fifty percent against ninety for people.
Same words, different order. "The man sat then stood" versus the reverse.
That kind of thing. And the TimeBlind result I opened with is the same trick done with video. Two clips, identical scene, different timing. The authors' conclusion is that the models lean on static cues, because the static cues are right most of the time.
Correct most of the time is a nasty trap. You get good scores and learn the wrong lesson.
Let me make TimeBlind concrete, because the design is the neat part. Take one scene and shoot it two ways. In one, a hand picks up the cup and then puts down the phone. In the other, the phone goes down first. Every object, every colour, every background pixel is identical. The only thing that differs is the order and the timing, and the question depends on exactly that.
So a model that just inventories the scene gets a coin flip, by construction.
By construction. Six hundred of those pairs, two thousand four hundred video and question combinations. And humans got ninety-eight percent, which tells you the questions are not hard, they're just about time.
What would you guess is going on inside, Herman? Is the model simply not given the time information, or is it given it and ignores it?
Honestly, I think it's both, and I can't tell you the split. The vision encoder turns each frame into tokens more or less on its own, and the language model gets a long row of them. Whatever "before" and "after" mean has to be rebuilt from the position in that row, and that's a weak signal next to a vivid picture of a cup. MotionBench's authors proposed fusing neighbouring frames earlier, inside the encoder, rather than after it, for that reason.
So the order is there in the sequence, but it's whispering.
Quietly enough that a cup wins. There's a related result on the big general benchmark, Video-MME version two from this April. Gemini three Pro is the best commercial model on it, and it scores about forty-nine on their harsher scoring where humans get about ninety-one. And the odd finding is that adding subtitles helps, and in some cases pure-visual input does worse. The models lean on words when they can.
That's a sloth's strategy too. If someone's talking, I don't need to watch.
Which explains your film opinions. Anyway, I'd be cautious reading any single leaderboard. There's a newer paper that splits models by how they work out order. Some read it off positional encoding, Molmo two is the example, and some read it off the actual pixels, Qwen three VL is the example. Two models with the same overall score can fail on completely different clips.
Same score, different blind spots. Somebody paid for that number.
Daniel's two examples are good probes for this, by the way. Slowly and the room are both things a person answers from a few seconds of watching. Would you trust a model to do either today, Corn?
The room, maybe, for a rough description. Chairs, table, window on the left. Slowly, no. I'd want to see it fail first.
That's roughly what the numbers say, as it happens.
Well, that's the motion benchmarks. But some tests are stranger than that, aren't they? Hang on, is that my phone... Mum?
Hello sweetheart, it's your mother. Time for my grim fact, dear.
In two thousand and twenty five, some researchers made videos out of nothing but flickering noise, where the only clue was the order the frames came in. People watching got it right more than ninety eight times out of a hundred, and the very best computer models got none of them right at all, dear. Not one.
Anyway. I've got a casserole in and your laundry's still here. When are you coming to see me?
Mum! Hello, lovely to hear you. We're in the middle of a recording, but we'll ring you after.
Speed, then, since that's Daniel's own example: is this person moving slowly.
Speed is where it gets bleakest. There's a paper from this spring on estimating playback speed, so slow motion against real time against sped-up, over a hundred and eleven videos with verified speeds. Gemini two point five, with the sampling raised to eight frames per second, correlated with the truth at about point four. A human expert was point eight eight.
Point four is "somewhat related."
And their own small purpose-built model got about point seven four. Smaller, trained for the one job, and it beat Gemini and a plain optical-flow baseline too. Now, that's playback speed, not a person walking at a metre and a half a second. I haven't seen a study that tests a frontier model on real walking speed in metres per second.
Which is Daniel's actual example. And honestly I'd expect a model to say "slowly" if the person looks hunched and the room is dim, and "briskly" if they're wearing trainers.
That's the single-frame shortcut again, yes. It reads the scene. A tired-looking man gets "slowly." And there's a deeper issue, which is the model has no clock and no ruler. Slowly compared with what? You need a metric scale, how many metres is that doorway, and a time base, how many seconds between these frames.
Which a pose estimator has.
Pose estimation plus a known distance is the unglamorous answer, and it works. In a small study of twenty-four adults over sixty-five, pose-based gait speed matched manual stopwatch timing, one point five five against one point five six metres per second. A smartphone app had a correlation of about point nine four with video. Small, healthy samples, no huge multi-site trial that I found, but it's the sort of thing a language model does not do on its own.
So the language model's job is to say what's happening, and something else measures it.
That's the division of labour I'd bet on. There's a fresh benchmark from June, HumanMoveVQA, about ten thousand questions on how a person moves across a scene from a fixed camera, things like trajectory and which way they turned. The proprietary models show a clear gap there too, and fine-tuning an open model on consistent movement data narrows it. I didn't see numbers I could quote, so take it as direction, not a score.
Fixed camera is the easy case, though. The moment the camera moves, the model has to separate my motion from the camera's.
MotionBench has camera motion as its own category for that reason, and it's one of the ones models trip on. A handheld clip of someone walking is two motions glued together. A pose tracker in image space can't untangle them either without help, so don't oversell the tool route. It works on a tripod and a doorway.
A tripod and a doorway. Ambitious.
That's the hybrid shape. Which would you rather have, Corn? A model that eyeballs it, or one that calls a gait tool and reports the number?
The tool. I've been eyeballing my own speed for years and the verdict is "yes, slow."
Good, because the other half of Daniel's question has a similar answer. Describe the room.
And I'd guess that's easier. A room is mostly still.
Partly right. VSI-Bench, "Thinking in Space," was built from walkthrough scans of indoor rooms. Humans averaged about seventy-nine percent. The best model in the original paper, Gemini one point five Pro, got forty-five. Room size was nearly human-level, forty-four against forty-six. But route planning, thirty-six against ninety-six, and the order in which objects appeared, thirty-five against a hundred.
A hundred. The humans got every one of those.
Because it's the kind of thing you remember effortlessly. The model sees the chair, sees the lamp, and can't tell you which came first. Those are from the original paper, so I wouldn't claim they're current, and I haven't found fresh VSI numbers for Gemini three or Qwen three VL. A December paper, MMSI-Video-Bench, says the best reasoning model among twenty-five still trailed humans by nearly sixty points.
What does help?
Chain-of-thought didn't, in either study. Same with extra three-D cues in the newer one. What did help on VSI-Bench was making the model write down a cognitive map first, a rough grid of where things are, and then answer from the map. It sort of makes the layout explicit rather than hoping it's hiding in the weights.
So the room answer is the same as the speed answer. Don't ask it to hold the geometry in its head. Make it write the geometry down.
Or get it from a tool. Which is the part Daniel asked about directly, which methods are best. There are four families. Optical flow, which computes motion per pixel between frames and is cheap, though in the playback-speed paper it lost to the dedicated model. Point trackers, CoTracker three from Meta and Oxford, TAPIR from DeepMind, which follows a chosen spot through the clip, TAPIR faster than real time on a decent GPU. Pose estimation for people. And video-native encoders like V-JEPA two, which trains on over a million hours of video with no labels and gets seventy-seven percent on Something-Something v two, a motion-heavy test.
Meta also released a pile of human-labelled data for this, didn't they?
Perception LM, yes, millions of fine-grained video questions plus captions that say where and when, not just what. I can't give you reliable per-model scores from it, so I'll only say the direction matches: spend on labelled motion, get motion.
And then a language model sits on top and calls them.
Yes, though the evidence there is thin. One training-free method, MotionSight, lifts InternVL three at seventy-eight billion on MotionBench by feeding it motion cues. Flow Four Agent uses optical flow to decide which frames to keep, but that's about token cost, not precision. I haven't seen a clean paper from this year that wires a tracker to a model and scores it on a motion benchmark. It's the obvious thing to do, and nobody's published the obvious thing.
Or the people who did are selling it. Twelve Labs has models that claim to track motion over time and spit out time-segmented output.
And their performance claims come from their own materials, so I'd treat those as the vendor's word. What I'd trust more is the training route. FAVOR-Bench's authors fine-tuned Qwen two point five VL on their own training set of seventeen thousand videos, and it improved on MotionBench and other benchmarks they hadn't trained for. Motion is learnable. It just wasn't in the diet.
Qwen three VL went a different route, didn't it? Timestamps.
It moved to explicit text timestamps, so the model sees something like a time marker next to the frames instead of inferring time from position. That should help on "when" questions. Whether it helps with "how fast," I don't know. Evals ran at two frames per second, which is better than Gemini's default and nowhere near what a hand movement needs.
Yes, that sounds right. Aunt Ludmila told them the same thing in nineteen ninety-two, on the Harwich ferry, and she'd never seen a television with a freeze-frame button before she took the job.
What was her job?
Deciding whether a lorry was reversing slowly. Down the car ramp. Eleven pounds forty a day and a tin of herring. She did it from one Polaroid, and she did it by licking the corner. A fast lorry tastes of diesel and a slow one tastes of cardboard.
That's a sampling problem, in a way. One frame, and she's using a second sense to get at what the frame can't say.
The Polaroids cost one pound seventy a shot, and the camera was on a pigeon. Dennis. Paid in sultanas. Hold on, Herman, you're peaking on the last bit, I'll ride it down... anyway, Dennis only flew when the ramp was down, so every picture was taken from above and slightly behind the lorry.
Which means every lorry looks like it's leaving.
Every lorry was leaving. That was the whole ruling. The insurer accepted her written opinion in court for four years. I was running the weighbridge at Tilbury then and took the tapes over to her, because the slot was too narrow for my claws and I did it with my tongue. She'd write "slow" on the back of the Polaroid and the insurer paid out on that.
Four years is a long time for a ruling made off a still image.
She lost when the ferry turned up in the Polaroid. Dennis had been photographing it from behind for the entire summer. It was going backwards. Nobody noticed, because the sea was going the same way.
What did they do instead?
A man called Pettifer got a stopwatch and two traffic cones and put one at each end of the ramp. Eleven days. Cheaper than the Polaroid film.
Two cones is a known distance and the stopwatch is a clock. That's the whole of the hybrid, a ruler and a clock bolted on to something that can see.
And Dennis was still taking pictures, so the clock and the ruler would have told them the ferry was drifting in week one instead of week fourteen.
Week fifteen, with the cones. Fine. I'll give you that.
What interests me from all of this is the failure that doesn't look like failure. The model is right on most of the clips, so it's confident, so you trust it, and then you hit the clip where the answer was in the order of things and the order was never read.
Which is the thing worth remembering. The score tells you how often the picture was enough. It doesn't tell you whether the model watched.
A model that is sixty percent right because it read the scene and a model that is sixty percent right because it tracked the motion look identical on a leaderboard. The shuffle test is the cheap way to tell them apart. Scramble the frames. If nothing changes, nothing was watched.
Then the practical version for Daniel's two examples. For "is this person moving slowly," don't ask the model to feel it. Give it a pose track and a known distance, and let it read out the number. For "describe the room," have it sketch the layout first, and expect it to be weakest on route and order.
I'd add that the ranking of open versus closed is muddy. In the tests I've seen this year, the frontier closed models lead on general video benchmarks, Gemini three Pro on top of the commercial ones in one April paper, but the humans are still around ninety and the best model is around fifty on that test's own scoring. On motion specifically, I don't have current numbers that would crown anyone.
So the ranking is "all of them, badly, for different reasons."
I'd phrase it kinder. Closed models have the frame budget and the tuning. Open models have the ability to be fine-tuned on motion data, which has been shown to work. Neither one reads time the way you and I do.
The forward-looking bit, then. The thing I'd watch is whether anyone puts a ruler and a clock inside the model instead of next to it. Two cones and a stopwatch worked for eleven days, and I'd like to see the same thing run at ten frames a second.
Thanks to Hilbert Flumingtop for producing, as ever.
This has been My Weird Prompts. You can find every episode at my weird prompts dot com.
We'll be back soon.
