It’s a typical evening. I’m at home, scarfing down dinner and taking a breather from an intense database systems cram session. My midterm exam’s in two days, but as always, my mind wanders to the realm of artificial intelligence. I recently stumbled upon an article by Yann LeCun, one of the big shots in AI. Now, LeCun’s an interesting character — he’s a leading figure in the field, but he’s also a vocal critic of the idea that current language models are on the fast track to general intelligence. His argument? Language itself is limited. LeCun’s got a point. Language, while grounded in reality, is essentially a human filter. We experience the world firsthand, then distill those experiences into words. It’s like we’re compressing raw data into a format that’s easier to transmit but loses some fidelity in the process. To put it another way, language is a projection of the real world, filtered through our human perception and cognition. We humans act as mediators, processing the raw reality and creating a distilled version of it through language. This realization highlights a crucial limitation of large language models: they process only this linguistic projection, not the real world itself. In a sense, they’re disconnected from reality, relying on our human-filtered version rather than direct experience. It’s as if they’re working with a shadow of the world, cast by human interpretation, rather than engaging with the world itself. This got me thinking: we’re really looking at two distinct types of AI here. On one hand, we’ve got AI that mimics human intelligence — which is what our current large language models are doing (in a limited way, agree that isn’t a perfect analogy). These models process the world’s web of text data and mimic it through next token prediction. It’s impressive, but it’s fundamentally an echo of human-generated content. This human-mimicking AI is about creating intelligence that’s similar to us humans — it thinks and outputs information in ways we’d find familiar. On the other hand, we’ve got this elusive concept of general intelligence — and this is where things get really interesting. We’re talking about artificial general intelligence (AGI) that doesn’t just mimic humans, but potentially surpasses us in intellectual capabilities. This kind of AI wouldn’t need to resemble human thinking to, say, cure cancer in the blink of an eye. It’s the real shit — intelligence that could solve problems in ways we can’t even imagine. The key difference? While human-mimicking AI is guided by our patterns and behaviors, true AGI would be unguided by human input. Instead of mimicking existing patterns, it learns by itself, more akin to reinforcement learning. This is the holy grail of AI — a system that can truly think and problem-solve independently, potentially far beyond our own capabilities. Creating human-mimicking AI? That seems almost straightforward when you break it down. I’m framing this as a supervised learning problem, you see. Think of humans as natural models, shaped by evolution instead of machine learning algorithms. We take in data through our five senses and output it through motor functions — essentially a five-input, one-output model. To create an AI version, you could theoretically slap a bunch of sensors on a robot (or a human in a fancy bodysuit), record all the inputs and outputs, and feed that data into a model for prediction. It’s classic supervised learning: you’ve got your input data (sensory information) and your output data (human actions or responses) to train the model. With enough of this labeled data, voila! You’ve got yourself a human-like intelligence that can predict outputs for new inputs, just like a human would. But here’s the kicker: do we really want AI that just mimics humans? What we’re after is that general intelligence — the useful kind that goes beyond human limitations. We want AI that can search problem spaces more effectively than we can. We’re talking about curing cancer, making groundbreaking scientific discoveries — finding those needles in the infinite haystack of valuable information, and doing it in ways that might be completely alien to human thought processes. How do we get there? For language models, supervised learning followed by reinforcement learning might do the trick. But what kind of data would we use? Right now, language is such a general form of data that it makes these models work. But for the real shit — the tough problems we want AGI to crack — I think we’re still in the dark. Language models have their place, but for those high-stakes tasks? We’re gonna need something more — something that transcends human-like intelligence and ventures into the realm of true artificial general intelligence. What that ‘something’ is, though, is anyone’s guess. We’ll just have to wait and see. Still developing my thoughts… trying to throw all the stuff on the wall and see what sticks.