There's a running joke in AI circles: if you can't explain what your model actually does, just slap "AGI" on it and watch the funding rounds materialize. Alexandre LeBrun, CEO of AMI Labs—the world model startup co-founded with Turing Award winner Yann LeCun—isn't playing that game.

"We never used the word AGI," LeBrun told TechCrunch in a recent interview. "And I just noticed that nobody is using it anymore; they switched to superintelligence. Next time we'll switch to something else." His take on the newest buzzword is equally unsparing: "There's no good definition. What is superintelligence? I don't know. It's not a very useful word."

That's a remarkably grounded stance for someone sitting at the epicenter of one of AI's most hyped new categories. LeBrun was speaking from Seoul, where he'd traveled for the International Conference on Machine Learning—and, more practically, to scout industrial partners in robotics, semiconductors, and manufacturing. AMI Labs is still pre-product. No demo, no deployments, no revenue. But LeBrun is already courting the companies that build things in the physical world, because that's precisely where he thinks the real problem lives.

What a World Model Actually Is (And Why It's Not Just a Fancier LLM)

Let's be precise about the technology, because this is where a lot of coverage goes soft. A large language model predicts the next token in a sequence—essentially, what word or phrase comes after what you've already written. A world model does something structurally different: it predicts the next state of a physical system. Push a glass toward the edge of a table, and you intuitively know it's going to tip and shatter on the floor. You didn't calculate fluid dynamics; you've internalized a model of how physical objects behave. That's the capability AMI is trying to encode.

LeBrun is careful not to position world models as LLM replacements. "Complementary, not replaceable," is how he frames it. Think of it like the brain's division of labor: one system handles language fluently, another handles spatial reasoning and physical intuition. LLMs remain the right tool for language tasks. World models are meant to fill the enormous gap those LLMs leave when they encounter anything with mass, velocity, or a floor to fall on.

That gap is enormous. "AI remains really dumb in the physical world," LeBrun said bluntly. Today's robots are largely running fixed, pre-programmed routines—deterministic loops that work beautifully until the environment changes by even a centimeter. The gap between a robot that can weld the same seam ten thousand times and one that can navigate a cluttered hospital corridor without hospitalizing anyone is not a small engineering footnote. It's the core unsolved problem.

The Safety Problem Nobody in Robotics Talks About Enough

LeBrun made a point that deserves more airtime than it typically gets: robots aren't safe right now. Not in open environments, not in homes, and definitely not around unpredictable humans. He cited a striking example—a robot at a public event that, lacking any contextual awareness, approached and kicked a child while doing a choreographed routine. The hardware executed perfectly. The "brain" had no idea a child was there.

"The hardware is very advanced; progress in hardware in the last few months is incredible, but there's no brain," LeBrun said. Context-awareness—knowing that a child is nearby and that kicking is bad—sounds trivially simple. In current robotics systems, it's an open research problem. Even achieving basic situational awareness would be, as LeBrun put it, "a very big difference for the world."

Scale that challenge up to a household robot, a delivery bot navigating sidewalks, or a surgical assistant in an OR, and you start to understand why AMI's pitch is about the physical world specifically. Factories with fixed routines are a solved (or at least solvable) problem. "Take your robot outside into a more open environment," and the entire engineering stack needs rethinking.

Healthcare: The 1% Problem

LeBrun has skin in this game beyond theoretical interest. His previous company, Nabla, was an AI health startup, so when he talks about medicine's limits, he's drawing on hard-won experience rather than thought experiments. His framing is pointed: today's LLMs in healthcare are like a doctor who aced every exam but never did a residency. Impressive on paper, dangerously incomplete in practice.

LLMs cover maybe "1% of healthcare," he said. The rest—physical examination, real-time patient monitoring, hands-on procedures, environmental reading—requires the kind of grounded, physical-world understanding that language models structurally cannot provide. World models, at least in theory, are the missing piece. Whether that theory survives contact with actual hospital workflows is another question entirely, and one AMI hasn't answered yet.

Why Asia, Why Now

Training a world model on the real world requires access to the real world. This sounds obvious, but it has enormous strategic implications for where AMI needs to operate. You can fine-tune an LLM on cloud GPUs in a San Francisco office. You cannot train a physical-world model without factories, robots, real environments, and partners willing to share them. "We need access to the real world, and it's easier for us to do that with partners," LeBrun explained.

That logic points east. South Korea specifically offers two things LeBrun considers rare in combination: a deep industrial base in exactly the hardware-heavy sectors that matter (robotics, semiconductors, advanced manufacturing) and a demonstrated cultural willingness to adopt technology fast. "Korea was the fastest adopter of the internet 25 years ago," he noted. That speed, layered on top of genuine industrial depth, is what he called "unique."

JP Lee, CEO of SBVA and one of AMI's Asian backers, framed it from the investor side. Korea's government has funded local sovereign language models that are "well enough" for general tasks, he told TechCrunch, but the real opportunity—and the one Seoul is now backing with a reported $880 billion commitment across chips, data centers, and physical AI—is in the physical layer. Lee argues that the language model problem is largely addressed locally; physical AI is where the next investment cycle should go.

Korea's developer community adds another variable: local engineers have a track record of not just adopting foreign tools but building on top of them to create homegrown platforms—Naver and Kakao being the canonical examples. For a pre-product company that needs real-world partners and fast feedback loops, that's a meaningful asset.

The Uncomfortable Truth About AMI's Current State

Here's where the skeptic's lens needs to stay on: AMI Labs has significant pedigree—Yann LeCun's involvement alone generates serious scientific credibility—and reportedly raised substantial capital. But it has nothing shipping yet. No product, no deployments, no public benchmarks. The company is, by LeBrun's own description, in an early scouting and partnership phase.

World models as a concept have enormous theoretical appeal. The argument that LLMs are fundamentally limited in physical-world reasoning is well-supported and not particularly controversial among researchers. But the distance between "we have a compelling research direction" and "we have a deployable system that makes robots contextually aware" is where most ambitious AI companies quietly dissolve. LeBrun refusing to use hype-laden labels like AGI or superintelligence is genuinely refreshing—but it also can't paper over the fact that the hard engineering work is still ahead.

The Seoul trip signals something real, though: if you're building physical AI, you cannot do it from a lab. You need the factories, the chips, and the robots. LeBrun is going to where those things actually exist. That's at least a sensible strategy, even if the destination is still a long way off.