Entrepreneurs Build 'Physical AI' World Models Globally

An Associated Press feature profiles a wave of startups and researchers redirecting attention from large language models toward so-called world models - AI systems designed to learn the statistical structure of physical space, time, and interaction rather than text alone. Computer scientist Louis Castricato left his doctoral program at Brown University to found Overworld (formerly Wayfarer Labs), which released Waypoint-1, a real-time interactive video diffusion model trained on 10,000 hours of video game footage, in January 2026 with a $4.5M pre-seed from Kindred Ventures. The AP frames the moment as a widening cohort of entrepreneurs and researchers drawn to embodied AI, even as investors continue to back language-model leaders such as Anthropic and OpenAI.

Where language models learn statistical patterns in text, world models aim to capture how light, motion, and physical interactions unfold across space and time. Overworld's Waypoint-1 is a frame-causal rectified flow transformer that runs on consumer hardware at 30-60 FPS and accepts real-time keyboard and mouse inputs - each frame generated using user controls as context, eliminating the latency common in earlier world models. The model was trained via diffusion forcing and post-trained with self-forcing to reduce error accumulation during long rollouts. The AP also profiles other researchers working on simulated training environments and real-world robot deployments as the field diversifies beyond text.

AI pioneer Fei-Fei Li describes "world model" as "one of the most important and most overloaded terms in AI today," according to the AP. Industry observers note the trend implies a practical shift from purely language-focused work toward embodied cognition, changing engineering priorities toward simulation fidelity, sensor integration, and datasets that capture physical dynamics over time - rather than large-scale text-model training alone.

The AP frames this moment as an extension of recent AI investment cycles. Language models delivered rapid advances in reasoning over text and unlocked broad applications; a subset of researchers and founders now pursue models that ground understanding in physics and action. For practitioners, the trend implies growing demand for tooling around sim-to-real transfer, robotics middleware, and benchmarks that test embodied generalization rather than text metrics.

Observers will track open-source world-model releases, embodied-task benchmarks, explicit physical-AI funding rounds, cloud-simulation partnerships, and early commercial deployments combining perception, control, and planning. Safety and governance developments as robotic and embodied systems enter public spaces will also be relevant.