opcodee

[opcodee @opc0de3]

TWO MEN IN A CARGO PLANE PUSHED A PALLET OF POTATOES OUT THE BACK RAMP INTO AN ACTIVE VOLCANO AND FISHED OUT CRISPS

396,000 likes. 1,397 comments

Nothing in it exists. No plane, no volcano run, no potatoes. It was generated, and the account behind it ships one of these a week

the money question is the interesting one, so start with what it did not earn:

→ instagram pays nothing per view. the reels play bonus was killed in march 2023, and adam mosseri said out loud they couldn't afford to run it in the us

→ so 396,000 likes converts to exactly $0 from the platform. that number is not revenue, it is a test result

→ what it proves is that the format holds attention for twenty seconds, and attention is the thing brands actually buy

→ rate cards sit around $400 a reel at 11k followers and scale with engagement, not follower count

→ and the line producing it costs a prompt and an afternoon, so margin per post is close to the entire fee

that last one is what people miss. a shoot like this was never expensive. it was impossible. a cargo plane, a volcano, an underwriter who signs off. no budget bought that at any price

so the business is not "post videos, get paid for views." that business closed three years ago. the business is running a zero-marginal-cost content line as a demand test, then selling the audience it proves exists

every category with a media budget is about to meet somebody doing this from a bedroom, with no crew, no permits, and no floor under their pricing

if you want to see how short that production line really is, run one still through image-to-video and describe only the motion. @Picsart does it from a phone

the platform stopped paying in 2023. the people making money noticed

[Lunar @LunarResearcher]

https://t.co/nN6fldtV4f

[opcodee @opc0de3]

A BLACK PANTHER SAT IN A PENTHOUSE LIBRARY, A LEOPARD WORE A DIAMOND COLLAR, AND A MONKEY POSED ON A LEOPARD-PRINT CUSHION FOR A FULL MINUTE

1M likes. 11,000 comments

It is an advert. The brand is a Dubai fashion label. In sixty seconds you do not see one item of clothing

think about what that minute actually bought:

→ no product, no price, no model wearing anything, no call to action beyond the name at the end

→ a menagerie of big cats in interiors that in reality means a location fee, a wrangler, insurance, and a permit nobody is granting

→ so the most expensive-looking asset in the entire feed cost almost nothing to produce

→ a million likes for a label selling clothes into a market of a few hundred thousand people

→ and the question nobody asks out loud: how much of that million is anywhere near Dubai

luxury advertising has always sold the world around the product rather than the product itself, and the world was the expensive half. a panther in a penthouse was the moat, because you either had the budget for that image or you visibly did not

that moat is gone. if anyone can generate the world, then owning the world proves nothing, and imagery that used to say "we are expensive" now says nothing at all

so the real question is not whether it worked — a million likes is a fact. it is whether reach acquired at zero marginal cost, from an audience with no ability to buy, is an asset or just a number shaped like one

if you want to see how cheap that world-building got, run one still through image-to-video and describe the motion. @Picsart does it from a phone

a million people watched a fashion brand show them no clothes. that is either the future or the entire problem

[Lunar @LunarResearcher]

https://t.co/nN6fldtV4f

[opcodee @opc0de3]

A WOMAN ON AN ELECTRIC SCOOTER, IN THE MIDDLE LANE OF A SIX-LANE HIGHWAY, KEEPING PACE WITH TRAFFIC

340,820 likes. Dashcam POV. Nothing about it looks staged

It is entirely generated. The creator says so himself, puts a label in the corner, and then asks the only question that matters: when did you notice, and what gave it away?

that question is the actual content. try it yourself before you scroll:

→ the traffic behaves. cars hold lanes, brake lights fire in the right order, the gaps open and close the way real traffic does
→ the dashboard reflection in the windscreen is consistent for the whole clip. that reflection is the thing your eye checks without being told to
→ the skyline sits at the correct distance and does not drift as the car moves
→ her posture is stable on a moving scooter at speed, which is the one thing a human would struggle to fake on camera
→ and it holds eleven seconds. the old tell was that these fell apart after five

the interesting part is not that it is convincing. plenty of things are convincing now. it is that the creator chose to label it and then turned the labelling into the hook

that is a genuinely new move. for two years the incentive was to let people believe it. he did the opposite, marked it, and got more engagement out of the reveal than the illusion would have earned on its own. honesty as a format, not as a compromise

if you want to calibrate your own eye, the fastest way is to make one yourself. image-to-video from a still, one line about the motion - @Picsart runs it from a phone. you start noticing the tells about ten minutes after you have made your own

he told everyone it was fake and it did 340,000 anyway. that should tell you where the attention actually comes from

[Lunar @LunarResearcher]

https://t.co/nN6fldtV4f

[Dmitro Cashback @DmitroCP]

THIS GAMEPLAY IS FROM GTA THAT DOES NOT EXIST YET. 96,157 PEOPLE LIKED IT ANYWAY

Minimap bottom left. Mission text top right. A chase through a city that has not shipped

None of it is a game. Nobody pressed a button. It is generated, frame by frame, to look like a screen recording

and the part worth understanding is which layer did the convincing:

→ not the street. a photoreal palm-lined road is table stakes now, every model does it
→ the HUD. the minimap, the objective text, the health bar — that is the layer your brain uses to classify "this is footage from a game"
→ the camera. third-person, slightly behind the shoulder, the exact rig every open-world game uses. it quotes a genre, not a place
→ the motion blur and the frame pacing. it stutters the way a console does, not the way a film does
→ and the cuts. store, bike, chase, escape. it has the rhythm of a mission because someone wrote it as one

so the tell is not realism. these things are past realism. the tell is that a generator reproduced an interface, and an interface is a promise about what is behind it. that promise is now the cheapest thing in the clip to fake

which is the new baseline for anything that claims to be "leaked footage" of anything. the pixels stopped being evidence about a year ago. the UI just stopped being evidence too

if you want to see how thin that layer really is, run one image-to-video test on a still with any interface drawn on it. @Picsart does it from a phone. the HUD survives the generation, and that is the whole trick.

the game is not out. the footage is. get used to that order

[opcodee @opc0de3]

FOUR CHARACTERS FROM A CHILDREN'S SHOW FOUND A BAG OF CASH, SMASHED AN ATM, AND RODE HORSES THROUGH A SUPERMARKET

175,584 likes

Cigar in the first shot. Stacks on the table. Then the spree

The clip is a joke about what money does to the innocent. The interesting money is somewhere else entirely

think about what that brand was worth, and why:

→ a children's character is an asset with one property that matters: control. every appearance is licensed, priced, and approved
→ the value is scarcity of use. the character shows up where the owner says, doing what the owner allows, and nowhere else
→ that scarcity is what a toy licence, a streaming deal and a theme-park contract are actually buying
→ and it just became impossible to enforce at the level of a single clip. anyone can cast the character in anything, in an evening, for cents

so the balance sheet question is not "can they sue." they can, and they will, one takedown at a time. the question is what a licensed appearance is worth when unlicensed appearances are infinite and better-distributed

the honest answer is that the legal right did not lose value. the practical exclusivity did. those were priced as one thing for forty years, and this year they came apart

which is the trade to watch. every rights holder is about to discover which half of their IP valuation was law and which half was just friction

if you want to see how thin that friction got, run one still through image-to-video and describe what should happen next. @Picsart does it from a phone. the character survives the render. the licence does not come with it

the characters found the money. the owners are about to find out where theirs went

[opcodee @opc0de3]

FOUR CHARACTERS FROM A CHILDREN'S SHOW FOUND A BAG OF CASH, SMASHED AN ATM, AND RODE HORSES THROUGH A SUPERMARKET

175,584 likes

Cigar in the first shot. Stacks on the table. Then the spree

The clip is a joke about what money does to the innocent. The interesting money is somewhere else entirely

think about what that brand was worth, and why:

→ a children's character is an asset with one property that matters: control. every appearance is licensed, priced, and approved
→ the value is scarcity of use. the character shows up where the owner says, doing what the owner allows, and nowhere else
→ that scarcity is what a toy licence, a streaming deal and a theme-park contract are actually buying
→ and it just became impossible to enforce at the level of a single clip. anyone can cast the character in anything, in an evening, for cents

so the balance sheet question is not "can they sue." they can, and they will, one takedown at a time. the question is what a licensed appearance is worth when unlicensed appearances are infinite and better-distributed

the honest answer is that the legal right did not lose value. the practical exclusivity did. those were priced as one thing for forty years, and this year they came apart

which is the trade to watch. every rights holder is about to discover which half of their IP valuation was law and which half was just friction

if you want to see how thin that friction got, run one still through image-to-video and describe what should happen next. @Picsart does it from a phone. the character survives the render. the licence does not come with it

the characters found the money. the owners are about to find out where theirs went

[Lunar @LunarResearcher]

https://t.co/nN6fldtV4f

[opcodee @opc0de3]

TWO CARTOON CHARACTERS SITTING IN THE BED OF A PICKUP UNDER THE MILKY WAY, WITH SNACKS AND A LANTERN

370,286 likes

Photoreal truck. Photoreal sky. Cartoon characters, rendered like they are physically in the frame

the production line this replaces:

→ a night shoot on a hill above a city, which means a generator, a crew and a permit
→ an astrophotography plate, which means a separate trip and a clear sky
→ a 3D team to model, rig, light and animate two characters
→ a compositor to marry the two plates so the lantern light lands on both correctly
→ weeks of scheduling, five figures minimum, and that is before anyone talks about rights

what it cost here: a prompt. the creator names the tool in his caption

that is the number that should be on every media company's desk. the expensive part of blended live-action and animation was never the idea. it was the pipeline, and the pipeline was the moat

but here is the line item nobody is pricing yet, and it is the one that decides whether any of this becomes a business: the characters in this clip belong to somebody. the generation cost nothing. the licence still costs what it always did, and no model output changes who owns the design

so the real lesson is split in two. production cost collapsed to near zero. rights cost did not move a cent, and it is now the entire barrier between a viral clip and a revenue line

if you want to test the production half on something you actually own, @Picsart does image-to-video from a phone - one still, one line of motion

the render got free. the permission did not

[opcodee @opc0de3]

TWO CARTOON CHARACTERS SITTING IN THE BED OF A PICKUP UNDER THE MILKY WAY, WITH SNACKS AND A LANTERN

370,286 likes

Photoreal truck. Photoreal sky. Cartoon characters, rendered like they are physically in the frame

the production line this replaces:

→ a night shoot on a hill above a city, which means a generator, a crew and a permit
→ an astrophotography plate, which means a separate trip and a clear sky
→ a 3D team to model, rig, light and animate two characters
→ a compositor to marry the two plates so the lantern light lands on both correctly
→ weeks of scheduling, five figures minimum, and that is before anyone talks about rights

what it cost here: a prompt. the creator names the tool in his caption

that is the number that should be on every media company's desk. the expensive part of blended live-action and animation was never the idea. it was the pipeline, and the pipeline was the moat

but here is the line item nobody is pricing yet, and it is the one that decides whether any of this becomes a business: the characters in this clip belong to somebody. the generation cost nothing. the licence still costs what it always did, and no model output changes who owns the design

so the real lesson is split in two. production cost collapsed to near zero. rights cost did not move a cent, and it is now the entire barrier between a viral clip and a revenue line

if you want to test the production half on something you actually own, @Picsart does image-to-video from a phone - one still, one line of motion

the render got free. the permission did not

[Lunar @LunarResearcher]

https://t.co/nN6fldtV4f

[opcodee @opc0de3]

@martynov014 hah yeah, strange woman

[opcodee @opc0de3]

SHE CLIMBED ONTO AN ESCALATOR HANDRAIL SEVERAL FLOORS UP IN A BANGKOK MALL AND LOOKED STRAIGHT DOWN.

36,337 likes.

Now price that. Because the piece I'm quoting does exactly that to the entire creator economy, and the arithmetic is uglier than the stunt.

Its author was driving delivery routes two years ago and now funds a startup on YouTube ad money. His numbers, on the page:

→ a Short of his did 22.9M views and paid €381
→ a 20-minute video did 247k views and paid $780

the view is not the product. the monetizable second is. and vertical video has almost none of them.

then the arbitrage almost nobody outside the US bothers to check:

→ "you can't monetize from here" is an old myth. ads serve to your foreign and VPN audience regardless and it counts
→ monetization connects through partner intermediaries who lend you their AdSense and pay to your card
→ RPM on local-language entertainment is low. on history or finance it is much higher
→ and in English, multiply the same view by 5 to 10

same video, same effort, different language, five to ten times the revenue per view. that is not a growth tactic, it is a pricing decision you make once.

the gate and the clock:

→ 1,000 subscribers and 4,000 watch hours in a year
→ that hours requirement moves to 8,000 in February 2027, so the cost of stalling is now on a timer

the dashboard he keeps, out of the thirty graphs YouTube shows you:

→ retention at 30 seconds. under 60% the video is dead, fix the opening. 70 to 80% and it has a shot
→ traffic source. mostly "Subscribers" means you are not growing. you want "Browse" and "Suggested" — that is the algorithm handing you strangers
→ click-through. 4 to 8% is normal for long video. under 3% and the thumbnail failed before anyone heard a word

on the cost side the build is an evening: script in Claude, voice in ElevenLabs, cut in CapCut, and the footage hour handed to @Picsart's Lina agent, which blocks out the scenes from the script. generated assets cost cents.

a clip like the one above risks something irreplaceable for 36,000 likes and a payout close to zero. the boring 20-minute video nobody screenshots is the one with a margin.

full playbook below ↓

[Kai Milstein @0xNotFounder]

https://t.co/sTmdpPh8Yh

[opcodee @opc0de3]

SHE CLIMBED ONTO AN ESCALATOR HANDRAIL SEVERAL FLOORS UP IN A BANGKOK MALL AND LOOKED STRAIGHT DOWN.

36,337 likes.

Now price that. Because the piece I'm quoting does exactly that to the entire creator economy, and the arithmetic is uglier than the stunt.

Its author was driving delivery routes two years ago and now funds a startup on YouTube ad money. His numbers, on the page:

→ a Short of his did 22.9M views and paid €381
→ a 20-minute video did 247k views and paid $780

the view is not the product. the monetizable second is. and vertical video has almost none of them.

then the arbitrage almost nobody outside the US bothers to check:

→ "you can't monetize from here" is an old myth. ads serve to your foreign and VPN audience regardless and it counts
→ monetization connects through partner intermediaries who lend you their AdSense and pay to your card
→ RPM on local-language entertainment is low. on history or finance it is much higher
→ and in English, multiply the same view by 5 to 10

same video, same effort, different language, five to ten times the revenue per view. that is not a growth tactic, it is a pricing decision you make once.

the gate and the clock:

→ 1,000 subscribers and 4,000 watch hours in a year
→ that hours requirement moves to 8,000 in February 2027, so the cost of stalling is now on a timer

the dashboard he keeps, out of the thirty graphs YouTube shows you:

→ retention at 30 seconds. under 60% the video is dead, fix the opening. 70 to 80% and it has a shot
→ traffic source. mostly "Subscribers" means you are not growing. you want "Browse" and "Suggested" — that is the algorithm handing you strangers
→ click-through. 4 to 8% is normal for long video. under 3% and the thumbnail failed before anyone heard a word

on the cost side the build is an evening: script in Claude, voice in ElevenLabs, cut in CapCut, and the footage hour handed to @Picsart's Lina agent, which blocks out the scenes from the script. generated assets cost cents.

a clip like the one above risks something irreplaceable for 36,000 likes and a payout close to zero. the boring 20-minute video nobody screenshots is the one with a margin.

full playbook below ↓

[opcodee @opc0de3]

THEY FLEW A PIZZA THE SIZE OF A CARGO RAMP OVER AN ACTIVE VOLCANO, COOKED IT IN THE CRATER, AND ATE IT ON THE WAY BACK.

159,336 likes. Posted yesterday.

Now price the real version of that shoot:

→ helicopter hire, with a permit to fly with the ramp open
→ a volcano authority signing off on filming above an active crater
→ aerial cinematographer, safety officer, and insurance on a crew working over lava
→ a food team building a prop pizza at that scale
→ and a weather window, which is the line item that actually kills the schedule

six figures, comfortably, and a month of coordination before anyone rolls a frame.

the version you just watched cost a prompt and some compute. the creator tags it as AI himself, so there is no deception in the clip. the deception would be pretending the old budget still means something.

what that does to the P&L of anyone who sells video:

→ the moat was never the idea. it was access to the helicopter
→ location, permits and insurance were the barrier, and those are precisely the line items that just went to zero
→ what survives is the part no budget ever bought: knowing that the argument about the crust at the end is funnier than the volcano

if you want to check it against your own numbers, price one shot you were quoted for last year, then build the generated version and compare. @Picsart does image-to-video off a phone — one still, one line about what moves in it.

the cost of spectacle collapsed. the cost of being worth watching did not move at all.

[Lunar @LunarResearcher]

https://t.co/nN6fldtV4f

[opcodee @opc0de3]

1,300 GPUs just beat a frontier model, and the number is the entire story

Liam Fedus built high-throughput materials labs in Menlo Park and last week posted what came out of them. the labs run experiments. the experiments make data nobody else owns. the models train on that data and then pick the next experiment. that loop is the product

his sentence, and the first two words are the ones to read twice: "using only 1,300 H200s, plus months of our experimental data, we mid-trained and RL'd an open-source model to surpass GPT-6 Astra on our analysis benchmark."

1,300 H200s is not a frontier run. it is a rounding error next to one. the footage here is his actual lab

what it implies if you are the one signing the compute invoice:

- the expensive part stopped being the training run. it is data nobody can buy, and here that data is manufactured by robots in a room
- they started from an open-source model. the base cost nothing. the money went into the half that could not be downloaded
- the target was their own analysis benchmark, not a public leaderboard. beating a general model on a narrow benchmark you own is a different and far cheaper problem
- the first domains are superconductors, magnets and semiconductor materials — all three are inputs to the compute supply chain that currently constrains everything else
- and it compounds. every experiment makes the model better at choosing the next experiment, which makes the next dataset better

the moat was never the model. it is the instrument that produces data the model has never seen

the catch is that this is one benchmark, defined by the people who beat it, with no third party reproducing it. read the number as a direction, not a result

whoever owns the experiment owns the dataset, and whoever owns the dataset stops needing the biggest cluster

[opcodee @opc0de3]

someone finally put a number on what AI does to a content business, and it's not the number you expect

in mid-2024 Pocket FM was at $200M ARR growing 50% a year. then Rohan Nayak pivoted the entire company to AI-only content production and growth flatlined for six months. he went on CNBC last week and gave the figures for what happened after that

audited FY26 revenue: $324M. current run rate: over $500M, up 70% year on year. content produced went from 25,000 hours a year to 2.5 million. and that audited year was their first full year of profitability, on EBITDA, on PAT and on free cash flow

the line he says almost as an aside is the whole story: they went from $250M to $500M and the cost didn't really go up

here is what changed underneath, with the numbers he put on camera:

- AI audio drama versus professionally produced audio drama is 80x cheaper. he says it twice and the anchor repeats it back to him
- 12-month revenue retention went from 44% to 76%. the pivot didn't just make content cheaper, it made it stickier, because more shots on goal means more hits
- 550,000 creators now produce 2.5M hours a year on their tools. they own the supply side and the demand side, which nobody selling generation tools does
- 90+ titles have cleared $1M in lifetime earnings and 13 have cleared $10M, out of a catalogue of 770,000
- Pocket Saga, the vertical-video version of the same catalogue, launched two months ago and is already around $15M ARR

the cost of making the thing fell 80x. the cost of knowing which thing to make fell by nothing at all. that's the moat, and it's made of 5 billion hours of minute-by-minute listening data

the catch is those six flat months. $200M ARR, growing 50%, and the pivot bought them half a year of nothing before it compounded. most boards reverse that decision in month three and never find out

"what used to take a big team can now be done by a single person." in this case that sentence is in the audited accounts

[opcodee @opc0de3]

two days ago a guy shipped a model that cannot write text. yesterday someone plugged it into a browser and searched flights for four tenths of a cent

$0.0039 per task. seven seconds. and the video is at 1x speed, he says so himself

Gregor Zunic, who runs Browser Use, wired his browser agent to Jev the day after it launched. 854,000 people watched the clip in a night

what he actually built:

→ a new action space computed at every step instead of one fixed toolbox
→ the page DOM handed over as state, not as a screenshot for a vision model to squint at
→ a small LLM kept in reserve, used only when something has to be typed
→ the whole thing open source under MIT, 811 stars on day one

why four tenths of a cent matters more than seven seconds

every browser agent today pays a frontier model to answer a question that is almost never a question. click this or that. is the page loaded. which of these nine buttons is the search box. that is a decision, not an essay, and you have been buying essays

when the decision layer costs $42 per billion tokens and the output is free, the cost of an agent step stops being the thing that limits how many steps you run

that is the whole unlock. not smarter agents. agents that can afford to be wrong twice

the sceptics in his replies are asking the right question: what did the old stack take. he has not published the side by side yet, and until he does this is one demo on one flight search

but the direction is not ambiguous. the people building on top of a two-day-old model are moving faster than the people arguing about whether it counts as a model

https://t.co/VTBwkO8gtW

clip is his, running at real speed

[opcodee @opc0de3]

a guy who helped build ChatGPT spent two years in stealth, then shipped a model that cannot write a single word

and the pricing is the part nobody is ready for

Diogo Almeida launched Jev yesterday. 17 million people saw the thread. almost nobody read the blog post underneath it, where the honest numbers live

what he actually shipped:

→ input costs $0.042 per million tokens. that is $42 per BILLION
→ output tokens are free. not discounted. free, because at this architecture they are too cheap to meter
→ end to end response in 70 to 500 milliseconds, against 3 to 329 seconds for a frontier model
→ the name comes from Jevons paradox. when a resource gets cheap enough, you do not save money, you buy vastly more of it

now the catch, which he put in his own thread: Jev cannot generate text. at all

it picks. it classifies. it decides. it returns a structured answer that matches a schema, so it cannot invent a field that does not exist

and here is the number to hold onto. the tweet says 20-200x faster and 40-400x cheaper. the blog says 40-200x faster, and then adds that their own benchmarks are "on the higher end of real world gains" because their team wrote the workflows

that line is worth more than the headline. a founder telling you his own number is flattering is rarer than a 400x claim

why this is a business story and not a model story:

every product that currently pays a frontier model to answer yes or no is paying string-generation prices for a decision. routing a ticket. flagging a transaction. checking if a document is complete. picking one of forty categories

those calls do not need prose. they need an answer and a confidence

if this holds up outside their own evals, the cost floor for that entire layer moves by two orders of magnitude, and the companies that built a margin on charging per decision have a problem

his own slide says it best: they are building Prod, not God

video is their launch reel

[opcodee @opc0de3]

@bcherny Tetris in Claude is such a perfect demo of how fast these models are getting at building things

[opcodee @opc0de3]

@martynov014 Build for where the models are going, not where they are today

[opcodee @opc0de3]

someone vibe-coded a Photoshop alternative with GPT-6 Astra, burned about $2,000 in tokens, and shipped it free. 170 users on day one

the math is the story. $2,000 is roughly two months of a junior designer's software budget at a small studio. it bought an entire image editor

what Photon actually ships with, from its own page:

→ layers, groups, masks and adjustment layers
→ smart objects and layer effects
→ subject selection and background removal
→ liquify, curves, command search
→ native PSD support, so files move in and out of the Adobe world
→ desktop app for macOS, Windows 11 and Linux

and the line that matters commercially: "Photon edits locally. There is no upload step, no queue, and no account needed"

so the pitch to a business isn't "cheaper Photoshop". it's: your client's unreleased artwork never leaves the machine. for agencies under NDA, that's the whole sale

Adobe charges per seat per month, forever. this was one person, one model, two thousand dollars, once

the uncomfortable question for anyone selling software: what does your product cost to rebuild now, and how long is your moat if the answer is a weekend and a token bill

[opcodee @opc0de3]

@martynov014 The scary part is that recursive self-improvement is no longer just a theoretical scenario

If multiple labs are seeing the same pattern, external evaluation becomes essential