Computer Use Is AGI for the Rest of Us — and GPT-6 Is Still Underrated
You thought the ceiling for LLMs was a slightly smarter chatbot? Bullshit. When GPT-6 hooks up to a robotic arm and starts painting Golden Gate Bridge oils, one after another, each one better than the last — that's when it clicked for me.
What does that mean? It means frontier AI has started operating the physical and digital world directly. And the most chilling part: the cost is falling off a cliff. A model at the Opus 5 tier just had its cost cut by 70%. You think that's a routine iteration? That's the formal launch of the kill program aimed at traditional white-collar work. By the time GPT-8 or GPT-9's Computer Use ships — even if the intelligence flatlines, as long as the price bottoms out — that's AGI for ordinary people. At that point it'll operate a computer better than you, and cost ten thousand times less. Is there still an ecological niche left for you in this world?
Your Iron Rice Bowl Is a Joke to It
We used to think Computer Use — letting AI control a computer — was pointless. Everything's going API-first, command-line-first now. Who still clicks a mouse? It's like hiring someone to bang on a typewriter after Word took over — isn't that taking your pants off just to fart? Then I saw a brutal real-world case, and I fully capitulated.
An indie developer who makes Mac cleanup software had AI download nearly four hundred common apps. Then the AI moved like a ghost, tracking which directory each app installed to, which folders were hiding junk. It didn't just analyze them one by one — it automatically extracted rules and pushed cleanup efficiency to an absurd level.
Ask yourself honestly: how long would that take a human? Hire an intern, have them download, install, and dig through folders every day — that's a shit-mountain you can never finish shoveling. But AI? Tens of hours without food or sleep, and it strips hundreds of apps down to their underwear. What does that tell you? Some purely mechanical work is garbage, but the data behind it is extremely valuable. Using cheap Computer Use to do this kind of dirty-data cleanup is a dimensional strike. You still think "experience" is worth something?
A Fruit Fly's 160,000 Neurons Rip Off Silicon Valley's Fig Leaf
Here's a more extreme example. Google just did something big: it digitized and open-sourced all 160,000 neurons of a fruit fly. A living brain, evolved over tens of millions of years, all in one file.
How long would it take humans to hand-draw those 160,000 neurons? Decades, and you still wouldn't finish. It's a dense three-dimensional microscopic plumbing network, ten thousand times more complex than a chip — and it's the size of a grain of salt. This time, AI did all the slicing, layer-by-layer scanning, and reconstruction.
What struck me most is that the underlying logic is identical to Computer Use. LLMs casually run into hundreds of billions of parameters, burn power like a small city, and still get crushed by a two- or three-year-old's intelligence. Why? Because LLMs are "brute force, big brick, just fly." A fruit fly has only 160,000 neurons, laughably low power draw, and it dodges predators and eats food. Isn't that a perfect "System 1"? Pure instinctive reaction.
Don't the people who spend all day bragging about parameter counts feel embarrassed? Using hundreds of billions of parameters to compute 1+1 — isn't that firing a cannon at a mosquito? The human brain runs on twenty or thirty watts, about the same as a bright lightbulb, and its intelligence crushes thousand-billion-parameter monsters. Same with Computer Use: the AI doesn't need to be a philosopher. It just needs to be like a fruit fly — see a button, click it; see data, move it. That's efficiency.
Silicon Valley's New Scam, Jev? Perfect for Dirty Work!
Everyone's hyping "Jev," the model from ex-OpenAI people. I just sneer. You call that innovation? Under the hood it's still a Transformer — essentially converting infinite generation into a finite multiple-choice problem. Isn't that just a product marketing show?
The lawyer in Breaking Bad who calls black white is the archetypal LLM — you have to feed it an ultra-long context and go back and forth several rounds before it gives you a conclusion. But Jev doesn't ramble. It only makes judgments and picks options, up to 400x faster.
I call it "Silicon Valley's new scam," but on Computer Use it's an absolute kill shot. Why? Clicking a mouse and moving data doesn't need long-form reasoning at all — it's a black-and-white multiple choice. Click this or that? Save or delete? Highly repetitive actions, done with this "fast thinking," and the cost drops straight to 0.5%.
Do enterprises adopt AI to trade poetry with you? Wrong. They adopt it to be a mover. A friend told me a company pushing "full-process AI-ification" — the boss's real KPI is cutting 30% of headcount every year. Enterprises have tons of data scattered across legacy systems or locked in employees' heads. Even more disgusting: an outsourcing team quotes 100,000 RMB to build a system but won't let you export the data — want the export? Another 100,000.
That's where Computer Use's value shows up. The boss just needs to take you to dinner, extract the process from your head, and have Computer Use record it and loop it in the cloud 100 times. Is that dirty work valuable? Valuable enough to cut your job entirely. One month of your salary buys decades of AI compute. What are you going to compete with?
The Human GUI Is Itself a Defective Product
What does the future actually look like? Definitely not a robot sitting in front of a screen clumsily operating Windows. People, you're taking humans way too seriously.
Things made for human eyes are themselves extremely inefficient defective products. We're visual animals — we need flashy UIs. A data flow that a few lines of code could explain has to be drawn as a picture before the boss understands it. Isn't that a bit pathetic?
The ultimate form of Computer Use doesn't need a monitor or a desktop OS at all. It becomes a headless Linux server in the cloud. Pure command line and API interfaces. Downloading and cleaning data, everything running at millisecond speed in a dark data center.
In the old days of digitalization, if you hit a vendor system that didn't play nice, just building the middleware converter ate your entire budget. And there were even more absurd approval chains: get a stamp from A, a signature from B, a review from C. In the future, these shit-mountains and absurd processes don't need to be refactored at all — just throw them at a cloud AI agent. It can walk through anti-human processes at a thousand times the speed. Isn't that better than spending ten million RMB rebuilding the system?
When the Price Breaks the Floor, the Singularity Arrives
Don't underestimate Computer Use's potential. When models drive the price into the ground, the Jevons paradox kicks in again — the cheaper it gets, the more people use it, and applications explode exponentially. Maybe in the next six to eighteen months, costs drop another order of magnitude, and the vast majority of mechanical labor in the world is wiped out.
We used to think AI was some lofty god. Actually it's just a worker who never gets tired, never takes a salary, and exists specifically to strip you down to your underwear.
Kind of darkly funny when you think about it: one day the company won't even have a living person who can sign off, and the AI will dutifully file its own request — then approve it with another model. That's what you call a cyber closed loop.
Don't think you're irreplaceable. Your experience is just a script to it.
[Subscribe to the JustSayAI Daily Brief] 📮JustSayAI Daily Brief · Subscriber Perk Two emails a day · One-click audio! Now fully available via email + IM + Skill.