Are LLMs a dead end?

We got intelligence, now what?

Those of us who’ve been in this corner for a while have watched these models improve almost in real time. As I’ve written before, I think today’s models can be as capable as most people at most knowledge work, if you give them the right setup: context, tools to call, the ability to steer them, and the judgment to know when to let them run and when to step in.

But… is that it?

Some of the people who built this field, such as Yann LeCun, and until recently Meta’s chief AI scientist, have argued for years that LLMs are a dead end on the road to AGI. He believes it strongly enough that he left Meta at the end of 2025 to start his own company built on a different bet entirely: world models.

His argument isn’t that LLMs are useless; on the contrary, he believes that they still are very useful in the short term. But, to achieve human-level intelligence, he argues that the architecture of predicting the next token, over and over, has a ceiling, and that no amount of scaling gets you past it.

Why next-token prediction might not be enough

So how far can a next-token engine actually get? Can you navigate the real world by predicting the next word in a sequence, or is the world simply too messy, too unpredictable, and with way too many random variables for that?

Take intuition, for example, human intuition is the output of millions of years of evolutionary trial and error; the pressure to survive turns out to be a remarkable engineer. Or emotions, which exist partly to remind us that actions have consequences, and to navigate the world, you have to act in it - and to act, something has to make you move.

What’s fascinating about LLMs to me is that the way they “think” is oddly human. We think largely by talking out loud or silently to ourselves. In an 1805 essay the German writer Heinrich von Kleist argued that we usually start with only a vague seed of a thought, and that the act of speaking is what turns the obscure into a whole idea . That’s not far from what a model does when it reasons step by step on the page.

But the big difference is that humans don’t need a prompt to act. We wake up with half-formed ideas and act on them. A model doesn’t, and for it to even be useful, one must push it in a direction. Sure, you can fake this with heartbeat cron-jobs and the likes, which make a model feel “alive”, but for us, the cron-job is innate: it fires the moment we open our eyes, because for us, acting means surviving.

If not LLMs, then what?

If LLMs aren’t the road to AGI, then what is? LeCun’s answer, and one shared by Google DeepMind’s Demis Hassabis, is world models: systems that learn how the world works from sensory experience, that understand cause and effect and physical dynamics rather than just the statistics of text. LeCun even prefers to drop the term AGI altogether - he calls the goal AMI, Advanced Machine Intelligence, because he doesn’t think human intelligence is all that “general” to begin with.

The other camp isn’t convinced, and plenty of serious people think LeCun is calling the game too early. They argue that scaling, better tools, and reinforcement learning on top of LLMs will close most of the gap long before any clean “world model” arrives, and that recent results keep moving the line (the case against the dead-end consensus). Even if they’re only half right, the practical answer to “are LLMs a dead end?” might be: it doesn’t matter yet.

Where I land

Which is roughly where I land: Maybe LLMs are a dead end for AGI, but that depends entirely on what you think the goal is. For me, personally, if this is the level of intelligence we get for the next few years, then I’m fine with that; it’s all the intelligence I need, really.

And notice that’s exactly LeCun’s own caveat: a dead end toward human-level AI, but very useful right now. Which is why the driver and its imagination matter more than ever. Give two people the same model, GPT-5.5 or Opus 4.7, and the same brief, and they won’t get the same result. If you haven’t decided what you’re building and why, the model will happily build something that sounds great on paper and that no one actually wants. Holding the system map in your head, the stack, the scope, what it should do, and, in my view, most importantly, knowing how to test it, is the real skill now. I’ve found my rhythm with these tools: knowing when to outsource the thinking and let them run, and when to step in, steer, and make the hard calls.

So maybe for once we could settle, consolidate, and integrate these models into how we work and into the economy, as well as stop obsessing over replacing human labour in favour of how these tools actually leverage us.

But that’s not how we work, is it? With that mindset, we’d never have made it this far. We always reach for a better future, the better tool - this is baked into our nature, it’s what makes us human.