This is the most copied walk on the internet right now. One blond bowl cut, one brown suit, zero self-awareness - and it lands every single time. Half the "sigma edits" in your feed are just this guy's confidence run through a filter. Feed his look into Grok and you'll have a hundred of him by morning. But the one thing no model can generate is the part that actually makes it work: a real person deciding he doesn't care what anyone thinks, and meaning it. Everyone's cloning the suit. Nobody can clone the audacity.

A year ago this cost $2M, a film crew, and a month in Morocco. Now it's one person, one prompt, one afternoon on a laptop. Cleopatra drifting a Lamborghini and dunking on a pharaoh - none of it is real, and your thumb still stopped. We're about to drown in footage of things that never happened. And this is the worst it will ever look again.

The funniest part of this clip isn't the dance. It's that he's performing the single most cloneable format on the internet in front of a nearly 400-year-old masterpiece that can never be cloned. Feed Grok "awkward guy dancing in an art museum, pink shirt and bowtie" and you'll have ten of him by tonight - but no prompt will ever reproduce the painting on the wall behind him. That one took a human a lifetime, and it's the only thing in the frame that was ever actually scarce. We spent 500 years making art impossible to copy, then built machines that make everything else free to copy overnight.

This is the most copied man on the internet right now, and every AI video tool is racing to clone him. You can already prompt @grok Imagine for "awkward confident guy dancing through a cheering office" and it'll hand you the suit, the bowl cut, the hallway. What it can't generate is the part that actually went viral: a real room full of real people choosing to hype one guy up. The format is infinitely cloneable - the moment isn't. That gap is the whole game in 2026.

You've seen this guy 40 times this week. By spring you won't be able to tell which ones were filmed and which ones were typed. The sad man in the cardigan isn't one video - he's a template. Retro color grade, one aching song, a caption that says "if I could go back." Emotionally legible, endlessly repeatable. That's the exact shape AI video was built to mass-produce. Type "1970s man crying in a car, golden hour, nostalgic" into Grok Imagine and it hands you a brand-new one in 20 seconds. Humans turned a feeling into a genre. The models are turning the genre into an assembly line. Read that again: the formats winning right now aren't the ones that are hard to make. They're the ones that are easy to clone. This isn't a trend going viral - it's a prompt going viral. The humans made the first one hurt. The hard part now is being the one still worth copying.

"Now everyone is building one" - here's one, days later. Not JeanPhil. His clone. Same DNA: an improbable, dead-serious man who shouldn't be winning, paired with a glamorous woman, in one absurd outfit, in a luxe room. New creator, new face, identical formula. All AI. this is the part the 100M-view number hides. what went viral was never the Frenchman - it was the archetype. the underdog who acts like the main character while the world plays along. that's a feeling, and feelings are copy-pasteable. so the clone doesn't need his face. it needs his posture: the bob, the moustache, the confidence that doesn't match the body. rebuild that emotional shape and you inherit the whole wave, minus the week it took him to find it. which is why "build your own JeanPhil" is already a dead end. the look goes free the second it works, so the look is worthless. the only thing that was ever scarce is being the one who invents the next feeling - not the one who reskins the last one. image-to-video off one still is how you test a character before the window shuts. he didn't copy a man. he copied a feeling. that's the only part that was ever worth stealing.

A FRENCHMAN WHO DOESN'T EXIST JUST DID ~100 MILLION VIEWS IN A WEEK - BLOND BOB, HANDLEBAR MOUSTACHE, HOUNDSTOOTH SUIT, THROWING SLOW AIR-PUNCHES DOWN A PARIS SIDEWALK His name is JeanPhil. There's no man in the suit. Every frame is AI - and this week there's already a $JEANPHIL coin and a creator "talking to his lawyers." A character went from a prompt to an IP in seven days. a thousand AI clips a day die at 2,000 views. why did this one break containment: → it reads from a thumbnail. the bob, the moustache, the houndstooth - you can ID him at 50 feet, muted, mid-scroll. that's the whole game on a feed → one signature move, repeated. the slow shadowbox. every clip is instantly "a JeanPhil," which is how a character becomes a format instead of a one-off → a catchphrase that costs nothing to reuse: "Oui Madame." the audience can quote it, which turns viewers into distributors → ordinary daylight, real Paris landmarks, handheld drift. it reads as filmed, so nobody drops into render-inspection mode → consistency is now just a reference image. same face, same suit, any street, any day - the thing that used to need the same actor, costume and makeup on every shoot here's the part worth keeping. the render was never the moat. JeanPhil isn't hard to generate - he's hard to invent. the scarce skill stopped being "can you make the shot" and became "can you design someone a hundred million people want to see a second time." that's a writing and casting problem, not a GPU one. and the tell is at the top: the moment it worked, it sprouted a token and a legal structure. the character outgrew the creator in a week. that's the new failure mode - not that you can't build one, but that you can't hold onto one once it's loose. image-to-video off one still is the cheapest way to test a character of your own. there's no man, no fight, no gym. there's a coin, a lawyer, and 100 million people who know his name.

THIS IS THE THIRD FRANCHISE IN 72 HOURS DROPPED INTO THE EXACT SAME TEMPLATE - DUMBLEDORE ON THE DECKS, HARRY ON THE KEYS, SAME WHITE VOID, SAME SUNGLASSES BIT Gandalf did it Saturday. Palpatine did it Sunday. Now it's Dumbledore. Same stage, same pose, same gag, new IP. All AI. the clip isn't the story, the template is. here's what's actually being cloned: → a white cyclorama - the one "set" that never has to be rendered twice. no environment to keep consistent across a dozen copies → two faces you already own: one revered elder, one young hero. the model doesn't invent anyone, it just has to not break a face Warner Bros spent 20 years installing in your head → the sunglasses-on-the-wizard beat, repeated verbatim - the single funniest frame from the first hit becomes the mandatory ingredient in every clone → a DJ/MC setup that needs almost no animation: a head nod, a hand on a key. the "performance" is three gestures on a loop now watch the decay. the first one (LOTR) did ~525K views. Star Wars, same template a day later, ~88K. this one's climbing, but it's not the first hit anymore. that's the full lifecycle of an AI-video template in one weekend - someone finds a shape the model nails cheaply, everyone swaps the IP, and the audience burns out on the shape before any single clip. the lesson for anyone making these: the template is free, which is exactly why it's worthless by day three. the model hands everyone the same trick at once. the only scarce thing left is being first - or changing the shape instead of swapping the skin. image-to-video off one still is how you test a format before the window closes. the render's clean. the idea's already been used twice. that's the trap of a free template.

GANDALF AND FRODO, DE-AGED AND DRESSED IN CHARACTER, DANCING LIKE A DJ DUO ON A BLANK WHITE STAGE 11,000 likes. 80 comments on half a million views. Ian McKellen and Elijah Wood's faces - down to the wrinkles and the curls - rendered mid-dance on a seamless white cyclorama. None of it was filmed. Every frame is generated. watch what ISN'T in the shot, because that's the whole trick: → there's no background. a plain white sweep is the single friendliest input a video model gets - no environment to keep consistent, no lighting that can drift, nothing behind them that can break between frames → the faces are borrowed, so the model never has to invent a character. you already know these two, so it only has to not break a face you remember → full-body dance is the flex, and the white void is what makes it survivable - no floor detail, no set, no continuity to betray the motion → 80 comments on 525,000 views is the tell. nobody argues with a clip they just watch. they don't stop to type, they loop it → the joke is pure recognition: two of the most solemn characters in film doing the least solemn thing possible. it lands before you ask how it was made here's the part worth keeping. this isn't a rendering achievement, it's a subtraction one. strip the background to nothing, rent two faces the audience already loves, and every hard problem in generated video disappears at once. what's left is a costume and a dance - the easy half. the hardest problem in AI video was always the world around the subject. put the subject on a white stage and there's no world to get wrong. if you want to see how far a plain backdrop carries a shot, image-to-video off one still is the cheapest test there is. @Picsart runs it from a phone. the faces are fake. the white room is doing more work than either of them.

A guy sprints across a quarry floor and dives headfirst into a hole in the rock - and the camera follows him all the way down, straight through a bottomless shaft until he hits water far below. One unbroken shot. None of it is real. the tell isn't the hole, it's the take. no drone, no diver, no rig threads a continuous plunge down a pitch-black well and comes out clean. a real crew would've had ten cuts hiding ten problems. the whole thrill here is that there's no cut to hide behind. starts as one still → image-to-video is where a shot like this begins. picsart runs it from a phone. there's no hole, no diver, no water. your stomach dropped anyway.

Five dump trucks back up to the rim of a live volcano and pour corn into it - the kernels hit the lava and pop into popcorn on the way down. None of it is real. It's the physics that gives it away: real popcorn needs even, contained heat, not an open lava lake that would incinerate it. But the scale and the oddly satisfying pour sell it before you do the math. starts as one still → image-to-video. @picsart runs it from a phone. the volcano's fake. the popcorn's fake. you still wanted some.

A fluffy white cat fishing in a tide pool while molten lava pours in three feet away. None of it is real. It reads like someone's nature footage - the steam, the little fish, the wet black rock - but lava, a calm cat and a fish pond don't share a frame in real life. Every second of it is generated. one still → image-to-video is where a shot like this starts. @picsart runs it from a phone. the fish are fake. so is the cat. you still watched it twice.

Would you hole up in a tunnel like this when it all goes sideways? Survivors seal a Colorado mountain pass — a truck and a wall of wrecked cars across the mouth, the dark end bricked off, a whole camp built in the gap. None of it was shot. Every frame is generated. a scene like this starts as one still → image-to-video. - runs it from a phone.

WALTER AND SKYLER FROM BREAKING BAD, REDRAWN LIKE A GTA LOADING SCREEN AND DROPPED THROUGH FIVE SCENES THAT HAVE NOTHING TO DO WITH EACH OTHER — AND THEY STAY THE SAME TWO PEOPLE THE WHOLE TIME 32K likes · 329 comments it's labeled AI, so no one's being fooled. what's worth taking apart is why this exact format — take a show everyone knows, redraw it, then put it somewhere insane — keeps hitting: → the faces are borrowed, and that's the trick. you already know Walt and Skyler, so the model never has to earn a character. it just has to not break one you remember → the GTA cel-shade is a cheat code for consistency. flat shading and hard outlines hide the exact texture wobble photoreal generation can't hold frame to frame → every scene is short and every cut is hard. nothing runs long enough to drift — the same move every weak render makes: leave before the seam shows → the setting does the comedy, not the animation. a gun standoff, a ball pit, a clinic doorway — familiar faces in the wrong place land before you ask how it was made → the voices are synthetic, but borrowed characters buy the cadence too. you fill the gap with the performance you already remember the format isn't a rendering flex, it's a recognition one. the model supplies the motion; you supply a cast the audience has watched for years. the hardest problem in generated video — making someone care about a face — gets skipped by renting a face they already care about. want to see how far a borrowed face stretches? image-to-video off one still is the cheapest test there is. @picsart runs it from a phone. the render is new. the characters did the work before you ever opened the app.

A HOUSE CAT IN A SILK SHIRT AND A GOLD CHAIN SQUARED UP TO THE CAMERA, RACKED A TOY SMG, AND POSED IN FRONT OF A WALL OF ITS OWN GUNS 88K likes · 2700 comments the clip is tagged AI, so there is no deception in it. what is worth taking apart is why the humanized-pet format works at all, because almost none of it is the animal: → the face is a real cat, kept almost untouched. that is the anchor. the eye locks onto a genuine animal face and forgives everything attached to it → the human body is where the generation happens, and it is the easy half. the shirt drapes, the shoulders sit still, the hands stay mostly out of frame. none of the hard animal motion has to be solved → the props do the character work. a chain, a shirt, a cartoon gun. each one is a costume cue that tells you who this cat is meant to be, and none of it has to render well, only read fast → the cuts are short. no shot runs long enough for the body to betray itself, which is the same move every weak render makes: leave before the seams show → and the attitude is the whole payload. a cat staring you down lands as funny before your brain finishes asking how the lesson under it: this genre is not a rendering achievement, it is a casting one. the model supplies a body, you supply the one real face a viewer cannot look away from, and the gap between those two is the entire joke the hard problem in generated video was always the performer. map a real animal's face onto a generated body and you delete it if you want to test how much the face carries, run the same body on two different pets. image-to-video off one still is the cheapest way. @picsart runs it from a phone the body is generated. the stare is real. that is the whole trick

Tavus gave me early access to Griffin, and I need to talk about the last 30 seconds of the call. Griffin is the first human interaction model: a fully duplex, video-to-video AI that watches and listens while it talks. It's the first model to pass the Turing test (45% of people thought it was real) and it ranks #1 on NVIDIA's independent Video Full-Duplex Benchmark, ahead of every Gemini, OpenAI and open-source system, 0.09 off a real human. None of that prepared me for the end. I'd mentioned early on that I was fighting a fever. Twenty minutes later, as we said goodbye, it stopped and said it hoped the fever broke soon. Unprompted. It remembered. It doesn't take turns and it doesn't wait for a prompt. It interrupted me, reacted in real time, and held one thread across the whole conversation. The rough edges are there if you look - a beat of latency, the odd expression that doesn't land. But it's the first time a screen felt like a person instead of a tool. And Tavus says this is the small model; they're not releasing the big one yet. Early access via @tavus · tavus.io/griffin · built by @hassaanrza and team.

nobody believes trump actually met aliens. that is not why this clip works. look at what the frame is doing, not what is in it: → it is shot through something. blurred gear in the foreground, a hard edge boxing both sides, like a lens poking out from behind a crate. that framing alone whispers "someone filmed this who wasn't supposed to" → the aliens are the least important part. what sells it is the boring stuff around them: the security detail standing slightly wrong, the tan vehicles, the aides in sunglasses. a mundane background is what makes an impossible foreground land → there is no clean face moment on trump. the whole thing stays at greeting distance, because a tight face is exactly where these still break → and it is one continuous shot. no cut to hide a seam the trick was never the alien. it was convincing you that you're looking at leaked footage instead of a production. that "leaked" look used to need a real long lens and a real vantage point. now it's a prompt. image-to-video off one still is the cheapest way to test a premise like this. @picsart does it from a phone. it is obviously fake. the interesting part is how hard your eye had to work to decide that.

A cat that doesn't exist just out-styled your last three posts. Bucket hat, gold chain, full swagger, every frame generated. The render was never the hard part. Knowing what's worth making is. @Picsart

This is the entire article in 15 seconds and nobody's face melted off. A miracle. Same two characters. Golden-hour romance, then rain-soaked money-in-the-car noir like they robbed a Saturday morning cartoon. Different world every cut. Same faces, same jackets, same humans. That is the ONLY hard problem in AI video, and it's the one everyone still fumbles. Prompt-only and you get a fresh stranger every scene, new nose, jacket gone, who is she. Lock the character as a reference and the model stops reinventing your cast every five seconds. Keep the character, change the world. That's the whole game. Ran mine through @Picsart. Zero faces harmed.

You proved it with a machine. I'll prove it with a prompt. Your snake wasn't real and it wasn't CGI. It was a 12-metre animatronic in real water, puppeteered on a stage. You built a whole machine, and the only reason to do that was her face reacting to something that actually moved. This clip is the opposite extreme. A door onto an impossible generated world, no stage, no rig, one prompt. Both routes fake the spectacle fine. Only one still can't fake the thing that matters: a real human reacting. That's the half that isn't free yet. Generate the world. Keep the human doing the looking. Made with @Picsart.