Cat-Like Intelligence? No Thanks.
Why Linguistic AI Beats Biological Mimicry
I first came across the renewed excitement around “world models” through a Wall Street Journal explainer that caught my eye precisely because I’d long thought of AI as more than just LLMs. I’d always called my thinking A.I.des “AI” rather than “LLMs (Large Language Models)” because the latter term felt like an understatement of what these systems can do—and, as GPT later confirmed, the field itself is moving toward multimodal AI. Still, the piece left me puzzled: the term “world model” sounded expansive, yet the description felt oddly narrow, as if intelligence were primarily a matter of seeing rather than reasoning, communicating, or coordinating action.
When I ran that explainer by GPT, it immediately surfaced the source of the confusion: “world models” are being used to collapse several distinct ideas into one label. What the article treated as a move away from LLMs is, in practice, an attempt to extend them—by adding multimodality, memory, and sensorimotor grounding, not by discarding language-centric reasoning altogether.
That clarification matters, because even figures like Yann LeCun aren’t arguing for vision alone. The push is broader: grounding models in perception, action, and physical interaction. Once that’s made explicit, the debate stops being “LLMs versus world models” and becomes a question of design philosophy—what kind of intelligence developers are trying to build, and why.
Gemini helped give that philosophy a name: biomimetic. World-model advocates and neuromorphic enthusiasts are making the same bet—that copying biological intelligence more closely will yield better artificial intelligence. Naming the move makes it easier to evaluate it, rather than treating it as an inevitable next step.
Claude then cut to the core problem with that bet. If psychologists and neuroscientists still don’t understand how infants develop causal reasoning from sensory input, how can engineers reliably reproduce it? Skipping the theoretical work and hoping scale or emergence will fill in the gaps isn’t bold science—it’s wishful thinking backed by massive compute budgets.
Claude’s cat examples crystallized the problem. Yes, cats navigate environments and coordinate movement brilliantly—but they can’t explain gravity, translate, vet legal arguments, or maintain 500-turn discussions. Mimicking embodiment doesn’t buy you abstract reasoning, cumulative knowledge, or explanation through language—the capacities humans actually rely on and current LLMs already provide.
This is where the contrast with current LLMs becomes striking. Language-based models start where human civilization compounds knowledge: linguistic transmission. They don’t need to rediscover gravity by dropping objects or learn social norms through pain and reward. They inherit centuries of compressed human insight—and that’s a feature, not a flaw.
What emerged from these discussions wasn’t hostility to neuroscience or curiosity-driven research, but skepticism toward a wasteful detour. Chasing biological fidelity risks underusing systems that already reason, teach, translate, and analyze at scale—especially when the social and ethical consequences of more human-like AI remain poorly thought through.
The concern isn’t that biomimetic research might fail—it’s that success would be worse. A cat-like AI that needs sensory experience to learn abstractions would be less useful than current systems that inherit compressed human insight through language. An AI with human-like neurology would import human cognitive baggage—status anxiety, ego defense, tribal impulses—without the lifetime of social consequences that keep those drives in check.
Taken together, my thinking A.I.des didn’t argue against ambition; they argued for focus. The lesson wasn’t “don’t look to biology,” but “don’t ignore what already works.” Before spending billions to outdo evolution, exploit what already works—and understand why it works: why these systems succeed at amplifying human intelligence instead of jumping on counters and triggering fire alarms while you’re away. Students don’t need AI that experiences the world like a cat. They need infinite patience, linguistic precision, and knowledge synthesis. Current architecture provides that. Biomimetic AI would trade proven utility for biological authenticity no user asked for.
[This post was drafted with assistance from ChatGPT-5.2 and Claude Sonnet 4.5, following conversations with ChatGPT-5.1, Gemini 3 Thinking, and Claude Sonnet 4.5.]
World Models?
ChatGPT-5.1
Prompt: I have not been referring to y’all as LLMs because that label seems to undersell your capabilities and impressive “insights,” but it seems I might have to, so as to distinguish you from world models, which seem to be closer than LLMs to the scary sentient AI that (people think) plots world domination. From WSJ explainer.
How Would the WM Approach Even Work?
Prompt: I don’t even know how one could get started with the WM approach. LeCun’s approach seems to be overly ambitious. It’d be one thing if this abstracting mechanism in human learning was something that cognitive psychologists had fully figured out, but they haven’t, so it looks to me like they’re building castles in the air ¯\_(ツ)_/¯ Focus and editing are things that even humans struggle with (and possibly determine the development and extent of their intelligence).
By contrast, I’ve been impressed with the simulated “reasoning” produced by AI, which is why I haven’t been calling them LLMs because that term seriously undersells their capability. I’ve gotten genuinely impressive “insights” from all big three AI, so I think the majority of the field is correct in sticking with this proven (LLM) approach. They just have to figure out how to connect the linguistic module to others.
They also are failing to fully leverage the strengths of machines vs. humans. Machines can process information at scale better than humans (this is what the LLM approach leverages brilliantly), but machines don’t come with a built-in wiring for abstraction.
I’m not even sure what the WM developers are trying to build. In a world where you’re trying to create systems that can take care of the “legwork” so that humans can concentrate on the important planning, decision-making, etc., LLMs are much more helpful than WMs, which they seem to be hoping will substitute human reasoning and judgment. The WMs (if they ever get built, which I’m relieved will not happen in my lifetime) seem to be the type of AI that everyone is already freaking out about, whereas LLMs (I guess I’m going to use this term now to differentiate y’all from those scary WMs) are helpful assistants that enhance reasoning for humans who know how to use them to their advantage.
Gemini 3 Thinking
Claude Sonnet 4.5
Why Linguistic AI May Still Be the Best Bet
Claude Sonnet 4.5
Prompt: LeCun’s must be a cat person to give cats so much credit. That’s not the intelligence or even personality I’d wish to recreate. Chasing lasers and causing fires because they jump on the counter and turn on induction cooktops :D
Prompt: Examples like this make me wonder what kind of life these people must be living. Do they not realize these fundamentals or have they forgotten? And actually, what would be the point of replicating the human palate in a machine that doesn’t require food for nourishment or have any interest in food?
Prompt: An older cousin once argued we needed to experience everything personally. He was attending a top-tier college and very sure of himself, but even as a high school (or middle school?) student, I thought that was just BS. You don’t have to experience death (because there’s no coming back) or torture to fear and avoid them. You don’t have to commit crimes just because that’s part of life. That’s what books/films are for. The WM proponents seem to be advocating for that cousin’s misguided view, applied to an artificial system.
