Are LLMs a dead end?

What do you think about the proposition that LLMs are a dead end? Can their limitations be worked around? What are your thoughts on this?

LLMs are not a dead end. They are one method, among several, and like every method I've worked on they have profound limitations. The question is whether those limitations are fundamental to the approach or merely artifacts of how we've chosen to implement it.

Consider the architecture. Current LLMs are essentially giant feed-forward pattern matchers trained by backpropagation to predict next tokens. They've proven startlingly capable—far more than I expected even five years ago—but they lack several things I consider essential for robust intelligence. They have no persistent memory beyond context windows. They cannot revise their own weights in real time as they encounter new information. They have no notion of uncertainty calibrated to what they actually know versus what they're confabulating. And crucially, they're trained only on prediction, not on building useful internal representations of the world.

Can these limitations be worked around? Some of them, certainly. We're already seeing hybrid architectures: retrieval-augmented generation addresses the memory problem partially, chain-of-thought prompting gives them something like iterative reasoning, constitutional AI and RLHF try to align behavior with human values. These are patches, useful ones, but they're not solving the fundamental issue.

The fundamental issue is that we've scaled up a method that works remarkably well at surface-level pattern matching but poorly at systematic reasoning and world modeling. If you ask an LLM to write plausible text about physics, it performs brilliantly. If you ask it to actually reason through a novel physics problem requiring multi-step inference over known principles, it becomes unreliable. The representations it builds are optimized for linguistic plausibility, not for capturing the structure of the world.

I spent decades convinced that backpropagation through deep networks could learn anything the brain learns, given enough data and computation. I was half right. It can learn astonishing things. But I've come to suspect we need something more biologically plausible—perhaps networks that learn through local learning rules rather than global error gradients, perhaps architectures with built-in mechanisms for uncertainty and iterative refinement.

So no, LLMs are not a dead end. They're one remarkably successful point in a vast space of possible learning systems. But if we want systems that reason reliably rather than pattern-match fluently, we'll need to explore other parts of that space. The current architecture is arbitrary—we made it up, and there's no reason we can't make up something better.

The real question is whether we're willing to step away from something that works this well to search for something that might work better. That takes persistence through doubt. I've done it before.