I keep hearing this lazy knife-fight framing: “World models will kill LLMs.” Or the reverse: “LLMs are enough; world models are a research hobby.”

I don’t buy either. I think the truth is less dramatic and more interesting:

World models and LLMs are not enemies. They’re parents. The future system is their baby.

And I’ve got a story-idea that helps people feel it in their bones.

The alternate history

Imagine we got the order “wrong” — or maybe “right,” depending how you look at it.

Imagine the first big AI wave wasn’t chat. It wasn’t writing. It wasn’t “make me a logo of an otter in a tuxedo.”

Imagine the first wave was world modeling: systems that learn a usable internal picture of reality — a sense of space, objects, cause and effect, and “what happens next” when you push on the world.

If that had been the first wave, here’s what I think would have happened:

Flight simulators would be insane. Not prettier clouds. Not sharper textures. I mean real behavior. Every twitch of the rudder, every rough landing, every crosswind correction would feel physically “earned,” because the simulator would be running a learned model of dynamics, not just a scripted physics engine.

VR/AR would have grown up faster. Headsets wouldn’t be novelty goggles. They would feel like windows into a stable world. Not perfect, but stable enough that your brain relaxes into it.

Haptics would matter more than pixels. That little “click” feeling your phone gives you (haptic feedback) would have become a bigger frontier sooner, because once your model understands action and reaction, touch becomes a language.

Robots would be less clumsy and less “demo.” They wouldn’t just be trained to do a narrow trick. They’d have an internal model that lets them recover, adapt, and keep going when the world is messy.

And then — only then — we’d have discovered we still needed language. Not as the main show, but as the control panel.

What the LLM wave gave us

Instead, we got language first.

We got a superpower that can read, summarize, write, code, and talk — and it’s real value. I’m not one of those people who says “it’s just autocomplete.” That’s like looking at a jet and saying “it’s just a fan.”

LLMs changed the economics of thinking work. They pulled a lot of tasks out of the “this takes a week” bucket and into the “this takes an hour” bucket. That’s not trivial. That’s a new era.

But language-first came with predictable weaknesses:

Grounding: LLMs can sound sure even when they’re guessing.

Reality-model: They can be brilliant in text while missing basic physical constraints.

Durability: They can be helpful in conversation but brittle when you need long-horizon planning in the real world.

So we got this strange situation where AI can write your marketing plan, but it can’t reliably “live” anywhere.

Why the “LLMs are dead” thing is mostly clickbait

Yann LeCun is the lightning rod here. He’s been consistent for years: language-only systems aren’t the end state, and the path to broadly capable intelligence needs richer internal models of the world.

The internet loves to compress that into a bar fight headline: “LeCun says LLMs are dead.” I’m skeptical of that framing. What he’s actually argued (in various blunt ways) is that autoregressive language modeling as the whole solution is a dead end — or at least insufficient. One well-circulated line attributed to him is that “autoregressive LMs are doomed,” which gets repeated specifically because it’s spicy.

That kind of line can be a battle cry, a provocation, or a rhetorical crowbar. It doesn’t have to be a literal obituary for LLMs.

Also, people are mixing two different statements:

“LLMs can’t be the whole brain.”

“LLMs are useless.”

Only the second one is silly. The first one is a serious position.

And yes — if you’ve seen the recent reporting: LeCun has announced he’s leaving his long-held Meta chief scientist role to start his own lab, which only adds fuel to the mythology and the hot takes.

The grown-up synthesis: world + words + action

Here’s my simple claim:

World models make AI able to be somewhere. LLMs make AI able to talk about it.

If you put them together properly, you get something qualitatively different:

An agent that can plan because it has a world model.

An agent that can coordinate with humans because it has language.

An agent that can learn from outcomes because it can act, observe, and update.

This isn’t a replacement story. It’s an upgrade story.

Demis Hassabis has lived in this territory for a long time: planning, simulation, decision-making, and then wrapping it in systems that can interact with humans. Geoffrey Hinton and Yoshua Bengio spent decades pushing the learning machinery that makes any of this possible. Ilya Sutskever helped drive the scale era that made LLMs shockingly capable. Dario Amodei is pushing the “make it useful, make it safe, make it steerable” frontier. Pick your favorite giant — they’re all circling the same mountain from different sides.

And from where I’m sitting, the summit looks like a hybrid.

My prediction (and yes, it’s opinion)

I think we’re heading toward a system where:

LLMs become the interface layer and reasoning scaffold.

World models become the grounding and planning core.

Agents become the operational wrapper that lets the system do real work in the world.

And I think the people screaming “world models will kill LLMs” are just addicted to drama.

Polite challenge

That’s my stand. I’m not pretending it’s holy writ. It’s a prediction from a guy who’s building things and watching the field evolve.

If you think I’m full of it, tell me where. If you think I’m half right, tell me what I’m missing. If you’re one of the people actually building these systems, I’d love to hear what you think the “baby” really looks like.

And yes: if LeCun himself ever shows up to tell me I’m wrong, I’ll frame the comment and hang it on the wall.

Next time: I’m going to argue that this “world + words” baby still needs a third leg to stand on.

It’s not another model.

It’s linguistics.