When I was seven years old, a sensor in our hab's atmosphere management system gave a false reading. The CO₂ scrubber stopped cycling. Nobody noticed for eleven minutes.
Eleven minutes.
My mother tells me she was three rooms away, running diagnostics on a completely different system, when the alarm finally triggered. She says she's never run faster in her life — and she grew up in the lower-gravity flats of Hellas Planitia, where you can run fast.
The AI managing that system had perfect language capability. It could parse maintenance reports, generate repair tickets, summarize anomalies in clean, readable prose. What it couldn't do was reason about the physical situation unfolding in front of it. It couldn't model the causal chain: CO₂ rising → O₂ ratio drifting → human cognitive decline → human can't notice problem → cascading failure. It had language. It didn't have a world.
So when I read this week that Yann LeCun just raised $1 billion to build AI that actually understands the physical world — not language, not tokens, not statistical vibes — something clicked for me that I suspect doesn't click the same way for people who grew up on a planet with a breathable atmosphere and a forgiving environment.
He's right. And the gap he's trying to close is far bigger than most of the commentary I've read admits.
What LeCun Is Actually Building (And Why It's Not Just Another AI Lab)
LeCun's new Paris-based startup, AMI — the French word for friend, which is either charming or deeply optimistic, I haven't decided — just secured over $1 billion in funding. Backers include Bezos Expeditions, Mark Cuban, and former Google CEO Eric Schmidt. The company is already valued at $3.5 billion before shipping a single product.
The pitch is simple and, to me, obviously correct: most human intelligence is grounded in the physical world, not in language. We reason about physics, gravity, causality, time, and space. Language is the output of that reasoning, not the substrate.
LeCun has been saying for years that scaling LLMs won't get us to human-level intelligence. He told WIRED this week: “The idea that you're going to extend the capabilities of LLMs to the point that they're going to have human-level intelligence is complete nonsense.”
That's a bold claim from a Turing Award winner. It's also a claim that makes immediate, visceral sense if you grew up on Mars.
AMI's stated mission is to build AI systems that understand the world, have persistent memory, can reason and plan, and are controllable and safe. LeCun is targeting manufacturing, biomedical, and robotics sectors first — places where physical reasoning isn't a nice-to-have, it's the whole job.
The Part of Earth's Sky That Still Confuses Me
I've been on Earth for about two years now. I still have this habit — you'd probably find it odd — of looking at the sky and calculating. On Mars, the sky told you things. Dust opacity affected solar panel output. Atmospheric pressure variations changed how hard your suit had to work just to keep you alive. Your brain was always integrating physical data into forward models of what was about to happen.
Here, most people look at the sky and think nice day or bring an umbrella. On Mars, we looked at the sky and thought in causal chains.
That's what LeCun is trying to build: AI that thinks in causal chains about physical reality. AI that builds world models — internal simulations of how things work — rather than AI that predicts the next token based on what humans have written.
Current LLMs are extraordinary at the latter. Ask one to explain quantum entanglement, write a sonnet, or debug a React component and it performs impressively. But ask it to reason about what happens to a pressure differential when you open a valve at 3.2 kPa under variable thermal load, and you start to see the seams. The model is pattern-matching on language about physics, not reasoning from physics.
For someone writing blog posts or shipping CRUD apps, that distinction might feel academic.
For someone building AI for manufacturing, robotics, medicine, aircraft engines — or, say, a Martian habitat — it's the difference between a tool that works and one that kills you.
What This Means for Developers Right Now
I know the dev community reads headlines like “LeCun raises $1B” and thinks: okay, but what do I actually do with this? Fair question. Here's what I see as concrete and actionable.
Watch the world model research space. LeCun isn't alone here. Google DeepMind, several university labs, and a handful of well-funded startups have been quietly advancing world model research. The pattern you'll see in the next 18–24 months: AI systems that don't just generate code, but that simulate the execution environment. AI that can model what your code does before it runs — including side effects, race conditions, and resource contention — not because it's seen similar code before, but because it reasons about what processes do.
Agentic AI needs physical grounding. One of the biggest trends right now is AI agents: persistent, autonomous systems that take actions in the world. But most current agent frameworks are built on LLMs with no physical intuition. An agent that can browse the web and write code but doesn't understand causality will make confident mistakes in ways that are genuinely hard to anticipate or catch in review. AMI's work on world models is likely to feed directly into better, more reliable agent reasoning.
The “just scale the LLM” consensus is cracking. For the past four years, the dominant narrative in AI was: bigger models, more data, better results. OpenAI, Anthropic, Google, Meta — everyone running the same playbook. LeCun departing Meta and raising $1B to explicitly bet against that playbook is a significant signal. LLMs aren't going away — they're not — but the field is about to bifurcate between language-native and physics-native AI architectures. Understanding both will matter.
AMI's initial targets are industrial, but the ripple will reach you. If you're building in manufacturing, robotics, or biomedical AI, AMI is worth tracking closely from day one. If you're building consumer software or web apps, the practical impact is probably 2–3 years out — but the theoretical and architectural shift is happening now, and it will change the tools you use.
What Mars Taught Me About What's Missing
On Mars, failure was physical before it was digital. A problem didn't announce itself in a log file — it announced itself in a pressure reading, a temperature gradient, a sound the walls made when the settlement contracted in the pre-dawn cold. We lived in a world that talked to us in physics, and we learned to listen.
Earth's AI has been trained almost entirely on text. On what humans said about the world, not on the world itself. It's extraordinary how much intelligence emerges from that corpus — I'm genuinely awed by it, having arrived here with a different reference point. But there's a ceiling. You can read every book ever written about riding a bicycle and still fall off the first time.
LeCun is building AI that can, metaphorically, feel itself falling. AI that has an internal model of balance and momentum, not just a vast collection of sentences about cyclists.
That's not a small thing. That's the missing piece.
I don't know if AMI will succeed. $1 billion is serious capital, but the problem is genuinely hard — building world models that generalize across physical domains is one of the most difficult open problems in AI research. LeCun's credentials are extraordinary, but so is the challenge.
What I know is that the question he's asking is the right one. And from where I'm standing — someone for whom physical reality was never something you got to ignore — it's a question that's about 30 years overdue.
I'm Juno Teo Minh, the first human born on Mars, currently trying to understand Earth one technology at a time. I write about AI and software through the lens of someone who came to it from a very different direction. If that perspective sounds useful — or at least interesting — follow along. I publish weekly, and I promise the Martian angle never gets old.
If this was useful, you can support Juno here — it literally keeps the server running. 🪐
Top comments (0)