I love AI. I hate AI. Usually within the same hour.
LLMs right now are really, really fucking good, and it looks like the labs have figured out how to keep making them better. The models knocked out my whole graveyard of half-finished projects in a weekend. Things I'd abandoned years ago, done, in two days. It's intoxicating. Letting Jesus take the wheel a little bit has me shipping easily 10x what I was before, and the high hasn't worn off.
That's also exactly what scares me. There's still a gap between "holy shit, I can do this now" and "holy shit, anyone can do this now." I just don't know how many model releases it has left. The thing that lets me ship 10x might mean nobody ever needs to pay me. What keeps me up at night isn't bio risks or data center buildouts. It's the profound sense of economic uncertainty. And it stings more because it's beating me at the one thing I'm actually better at than most people: breadth. I've always been the guy who read a lot, but the model has read the internet.
AGI is always n+1 capabilities away
So is it AGI yet? The answer is always "not quite," in two ways.
In people's heads, every new capability stops counting the moment a model has it. In the labs, the current RL paradigm means each new capability seemingly needs its own training environment. Spreadsheets? Build an environment. Browsers? Build another. That's not one general mind waking up.
But does that matter? Economically useful tasks are a tiny subset of all tasks. If you can saturate that subset one environment at a time, the economy doesn't care whether the thing doing it is "truly general." Neither does my landlord.
We've been moving these goalposts forever. Aristotle argued that reasoning was what made us special, until we discovered we're not actually very good at it and computers proved better than us at formal logic. Then it was the Turing test. That one fell quietly when we decided it had been a bad test all along. For a while humans were "just biological computers," until actual computers got better at computing and we realized that wasn't it either. Soon enough, the only thing left defining AGI will be whether it has a soul. Conveniently unmeasurable.
Part of the confusion is that "AGI" usually just means "human-like," and humans aren't general. We're extremely specialized: great at faces, shitposting, and throwing things; terrible at arithmetic, probability, and holding more than a few things in our heads. We call ourselves general because we're the reference point.
As most people use it, the term is so ill-defined it's useless. Nobody means the Legg and Hutter version, and anyways that's uncomputable. What "AGI" means is a vibe, and by that standard, VGI is already here.
Vibes all the way down
Which is the real problem: models are close to impossible to evaluate. The evals lie, often. Benchmarks saturate and get gamed, so we have to run on vibes. Someone posts a jaw-dropping demo on Twitter, a counter-thread shows the model insisting there are two r's in strawberry, and a consensus forms for about 72 hours until the next release.
I'm not above it. My eval is vibes too: does it get my project over the line? It's the only benchmark I trust, and the current models have me dreaming 100x bigger.
So I'll keep shipping. But the next time someone starts in about LLMs and AGI, I'll probably just start arguing definitions to bury that worry down. While we argue about the word, the economically useful tasks are getting eaten one RL environment at a time. The better question is what it can do this month that it couldn't last month, and what that means for me.
Sources
- Aristotle, Nicomachean Ethics, Book I, Chapter 7. Perseus Digital Library.
- Wason, P. C., "Reasoning about a rule", Quarterly Journal of Experimental Psychology 20(3), 273–281 (1968).
- Portoraro, F., "Automated Reasoning", Stanford Encyclopedia of Philosophy.
- Jones, C. R. & Bergen, B. K., "Large language models pass a standard three-party Turing test", Proceedings of the National Academy of Sciences 123(21) (2026).
- Legg, S. & Hutter, M., "Universal Intelligence: A Definition of Machine Intelligence", arXiv:0712.3329 (2007).