Overworld founder pivots from chatbots to world models in AI

Louis Castricato spent the better part of a decade working on the kind of AI that powers ChatGPT and Claude. Then he decided the interesting problems were somewhere else entirely.

The computer scientist, who built deep expertise in large language models and reinforcement learning from human feedback, launched Overworld (formerly Wayfarer Labs) with a straightforward thesis: the era of fundamental LLM breakthroughs is winding down, and the real action is shifting to what the industry calls “world models,” AI systems designed to interact with physical environments in real time rather than just generate text.

From text generation to physical simulation

“We basically have passed the point of doing real fundamental LLM research. Now it’s just applications.”

That’s Castricato explaining why he walked away from the LLM world. His argument isn’t that LLMs are useless. It’s that they’ve entered the “build apps on top” phase rather than the “discover new paradigms” phase.

Overworld’s answer to that paradigm gap is a category of AI called diffusion world models, designed to run locally on consumer hardware. The company’s Waypoint-1 model, unveiled in January 2026, targets 60 frames per second with less than 20 milliseconds of latency. In English: it aims to generate interactive, game-like environments fast enough that you wouldn’t notice the AI is improvising every frame.

That’s a meaningful technical bar. For reference, most competitive gamers consider anything above 30ms latency to be noticeable. Overworld is targeting sub-20ms, which would put its AI-generated worlds in the same responsiveness tier as traditionally coded video games.

The team and the money

Castricato isn’t building this alone. Overworld is co-led by Shahbuland Matiana as Chief Science Officer, and the founding team includes alumni from Stability AI and CarperAI. The company is based in Providence, Rhode Island.

Overworld closed a $4.5 million pre-seed round in January 2026. Kindred Ventures led the deal, with participation from Amplify.LA and Garage Capital.

That local-first approach is worth unpacking. Most cutting-edge AI models today require cloud servers with specialized chips to run. Overworld wants its models executing directly on the devices people already own, which removes latency penalties that come with round-tripping data to a remote server.

Why this matters beyond gaming

The obvious application is video games. An AI that can generate interactive 3D environments at 60 FPS could fundamentally change how games are built, shifting from hand-crafted levels to procedurally imagined worlds that respond to player behavior in real time.

But Overworld is explicitly positioning itself for broader use cases: robotics, physical AI simulations, and what the industry increasingly calls “embodied AI.” The logic is straightforward. If you can teach an AI to model a physical world convincingly enough for a human player to navigate, you can probably teach it to model a physical world convincingly enough for a robot to learn from.

One important caveat: Overworld has explicitly stated it has no connection to any cryptocurrency or token project. The company is focused on hardware-based AI systems, full stop.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.