We Do NOT Have AGI…Yet
Everybody and their dog has an opinion on AGI right now, and frankly a lot of those opinions are shit. Not because the people are dumb (well, yes, quite a few are), but more so because their definitions suck and they’re ass at making distinctions.
Some set the bar at “passes a Turing test,” which is hilariously low. Others demand consciousness, which is fucking stupid.
As far as I’m concerned the folks claiming current LLMs are already AGI are either retarded, or selling something.
And no, I give not a single fuck if the person who invented the precise term thinks we have AGI. He’s wrong, and clearly has a shitty definition 🖕🏻
So for the record (again) here’s my definition, the only one that really earns the G in AGI.
AGI is a system that:
Has genuine understanding. Not rote memorization. Not statistical correlation over token sequences. Actual understanding of how things work, such that it can generalize what it knows to new contexts instantly, the way you and I do.
Is capable of making intuitive leaps to solve problems and accomplish goals *outside its training data.* Can connect dots that aren’t otherwise explicitly connected. Can synthesize, not just interpolate.
Can do, or is capable of learning to do, anything an average human can learn to do, at least as well as the average human (and ideally in the top decile or top percentile, since average humans kinda suck). Mentally if disembodied, physically as well if embodied. Obviously this includes the need for and ability to learn continuously.
All three are necessary, and none is sufficient alone.
Also while we’re at it, I might as well include my definition of intelligence:
Intelligence = predictive accuracy. Full stop.
Yes, there are other things that support that function, but EVERY one of those things is directly in service to the predictive accuracy.
Moving on.
Understanding ≠ pattern matching.
When object permanence finally clicks into place as a kid, that understanding transfers everywhere, instantly. You don’t have to re-learn it for every new object.
That’s what understanding IS. You learn a principle, you apply it somewhere you’ve never applied it before, and it works because you understand the principle. One-shot.
LLMs don’t have this. GANs definitely don’t have this (though they have gotten better).
A system that actually understands object permanence would nail it in a completely new context it had never been trained on. The generalization would be instant, not statistical fetching.
They could handle occlusion like a boss, which even the best GAN still doesn’t handle perfectly and consistently.
And this makes the “stochastic parrot” question actually testable…can the system get it right the first time in a format it’s never seen? One-shot cross-domain transfer?
Intuitive leaps are NOT the same thing as understanding.
Understanding is about generalizing what you already know. Intuitive leaps are about making or arriving at something new. Synthesis.
You pull together stuff from totally unrelated places. Sensory experience, half-remembered facts, vague pattern fragments, old conversations you forgot you had. And out of all that noise you synthesize something that didn’t exist in any of the inputs. More than the sum of the parts.
Darwin looking at finch beaks, geological strata, livestock breeding, and Malthus, then arriving at natural selection. None of those inputs contain the output. His brain MADE that connection, it wasn’t stored in his weights.
https://youtu.be/X9RYuvPCQUA
LLMs can do something that looks like this, but when an LLM connects dots, those dots were almost always already connected somewhere in the training data.
It found the connection, it didn’t make it.
A system that only finds connections is bounded by its training data no matter how big the dataset gets. A system that makes connections is bounded only by what it can perceive and what it’s willing to imagine.
Generality is what puts the G in AGI.
A system that understands and synthesizes but only in narrow domains is just a very impressive narrow AI. The system has to be able to continuously learn and perform across the full range of things a human can learn, not just the things it was trained on.
If it’s disembodied, it should match or exceed human cognitive performance. If it’s embodied (think robots), physical performance too. Most definitions either demand physical capability (which conflates intelligence with robotics) or ignore the body entirely.
And the “top decile or top percentile” part matters more than you might think. A system that barely clears “average human” on most tasks would technically pass a weak AGI test, but come on, nobody would look at that and feel like the word “general” had been earned.
Average is the floor.
So where are we?
NOT FUCKING THERE YET 🤣
I know that annoys people. There are AI researchers and tech CEOs and PLENTY of unhinged randos on Twitter and Reddit who will tell you with a straight face that we’re basically there. That the next model will cross the line, for real this time.
And the demos ARE impressive, I’m not going to pretend they’re not. Capabilities have improved a lot. I’m impressed both with where we are, and with how far and how fast we got here.
But impressive is not the same thing as general. A calculator is impressive if you’ve never seen one. Doesn’t mean it *understands* math, it just *computes* it.
A close-up magician is impressive if you don’t know the gimmick. Doesn’t mean it’s actual magic 🤣
LLMs are glorified Chinese rooms. I know that pisses people off, but the Chinese Room gets dismissed way too fast because everyone focuses on the wrong part. The standard rebuttal is “but the room as a system understands.”
No it doesn’t. The room FUNCTIONS. It doesn’t fucking understand. Maybe to a functionalist that’s sufficient, but that’s also retarded.
An LLM is a processor, not a synthesizer.
A processor transforms inputs to outputs according to rules. You can make it arbitrarily complex, doesn’t matter, still a processor. You can make the room bigger, hire more people to shuffle symbols faster, add more filing cabinets full of rules. None of that turns processing into understanding or synthesis.
A synthesizer produces outputs that aren’t in the rules or the inputs. Something new shows up, out of distribution.
LLMs are wildly sophisticated processors, and the average human is stupid and/or gullible, and that is exactly why the debate won’t die. When your processor is trained on a trillion tokens and can interpolate across that space with high fidelity, the output LOOKS like synthesis if you don’t know any better.
But it’s not.
The “intuition” is a statistical ghost. And ghosts, like sorcerers, aren’t real no matter how good the special effects are.
**What’s actually missing?**
People in the field are already converging on three things:
World models.
A persistent, bounded self.
Continuous learning.
You can’t generalize across domains without an internal representation of how things actually work. That’s what a world model gives you. An LLM has no world model. It has statistical correlations that sometimes mimic one, it has shadows of our own world model due to our language encoding it, but a system with a real world model can reason about scenarios it’s never seen. It can realistically simulate. Big difference between a system that can talk about physics and one that actually understands and reasons using physics.
Intuitive leaps need a perspective, a place where inputs accumulate and collide over time. That’s what a bounded, persistent self gives you. Without it, there’s no half-formed ideas sitting in the background waiting to smash into something new, and no place in the world.
You know how sometimes you wake up (or shower, or go for a walk) and the answer to a problem just… arrives? Your brain was working on it behind the scenes. An LLM can’t do that (agents are getting closer, but they are not yet fully persistent, and context is limited). Every conversation is stateless. Nothing is brewing. Nothing CAN brew. The pot resets to empty every single time (setting aside LLM “memory,” which is not really memory but a context and retrieval hack.
And a system frozen at training time is bounded by its training distribution forever. Doesn’t matter how large it is.
Continuous learning is what lets a system expand its own capability frontier in real time. You and I don’t stop learning when we leave school. We don’t stop learning EVER. An AGI by definition can’t stop either.
These aren’t independent problems, by the way. A world model without persistence can’t accumulate refinements. Persistence without continuous learning is just a static memory bank. Continuous learning without a world model is just appending data with no structure to hang it on. Pull any one out and the other two fall apart. You need the whole stack or you don’t have AGI. You just have a really good tool.
**So no. We do NOT have AGI.**
We have very impressive processors that fake understanding and mimic synthesis through sheer scale and linguistic relational correlation. The coverage is broad enough that the gaps are hard to spot in casual conversation.
But the gaps are real, and more parameters alone won’t fully close them. More data won’t fully close them. More scale will get you a better processor, but it won’t get you a genuine synthesizer.
I still think we’ll get there, and soon (within the next ~year), but nothing in the public eye today truly earns the G in AGI.
Ignore anyone who says otherwise.