An OpenAI engineer posted a clip last week of Astra beating all 48 levels of an "I'm not a robot" puzzle game. No mistakes. The irony did the work, a machine walking through the test designed to keep machines out, and it was everywhere by lunchtime.
Good clip. Honestly, I laughed. It's quirky watching it work, almost human eh.
But don't miss what it actually was. Astra finished a puzzle game about CAPTCHAs. It didn't crack a real security system on a real website. No one independent watched the run, the guy didn't share how he did it, so nobody can repeat it. "AI completes a game" is the honest headline. "AI beats CAPTCHA" is the one that went round. Those aren't the same thing, and the second one is a lot more exciting than what happened.
That gap? That's the whole launch.
OpenAI didn't say "here's a good new model." Greg Brockman stood up on the 3rd and called Astra the most intelligent and aligned model yet, then signed off with "welcome to the AGI era." Jensen Huang at Nvidia, the man selling the chips it all runs on, said "AGI has arrived." Massive words. And then you go and actually look at the demos everyone's waving around as proof, and they get thinner the harder you stare.
Somebody bothered to check the best ones. About a fifth of the strongest launch-week demos had never been reproduced by anyone independent. Posted, shared, believed, that's the whole chain. There was a viral thread swearing Astra had beaten Portal, grown a forest of four thousand trees and written a Bach chorale. Nobody could verify a word of it. Didn't matter. It flew anyway.
There's no proof Astra is AGI. I'll just say it flatly because nobody in the industry will. General intelligence means a machine that can do basically any thinking task a person can. Not one that's brilliant at 3D models and code and then invents a fact the second you actually depend on it. The benchmarks have Astra about level with Claude's best, not miles ahead. Two labs, neck and neck, one of them with a brass band. That's the real picture.
And look, the marketing I could shrug off. It's the next bit that gets me.
You keep telling people this is AGI. They use Astra. It's good, but it still forgets what they said two minutes ago, still makes stuff up, still needs checking, same as every model before it. So now you've taught the entire public what "AGI" means. And you've set the bar down on the floor. The scariest, heaviest word in this whole field, and millions of people now hear it and picture a chatbot that helped them write an email.
So what do we do when the real thing turns up? When something actually general arrives, the kind of thing we probably should be a bit frightened of, the kind that should make everyone stop and ask who's actually in control of it? Nobody blinks. "AGI? Yeah, had that for years, it books my tennis court." The word's gone. Spent. Worn out on a puzzle-game clip and a 3D Tesla that comes apart into little pieces.
That's crying wolf, and it's not a harmless bit of fun. We're spending down the one word that's supposed to make people pay attention, just to put some shine on a launch.
I'm not the AI grump here, by the way. We use this stuff daily. We build on it. I think it's remarkable, which is exactly why the loose talk winds me up. The people who get this best should be the ones keeping the words honest, and instead the loudest ones in the room have a model to sell or a warehouse of chips to move.
For you, running a business? Almost boring. Nothing changed on the 3rd. New model, bit better at some things, it'll show up in your tools soon enough. Same three questions as always. Does it do a job you care about better than last month. Does it fit what you've got. What's it cost. "Is it AGI" isn't on that list and never was.
But the day it does matter, I'd quite like us to still have a word for it.