World Labs CEO Fei-Fei Li has unveiled Atlas, the company's new world model designed to understand and generate 3D environments. The technology has potential applications in robotics, design and science. Li highlighted spatial intelligence as a significant frontier beyond language models. On AI safety, she emphasised that humans must maintain control over systems. She advocated for collaboration between independent benchmarks, academia, government and industry in evaluating increasingly powerful AI systems. The AI pioneer discussed these developments during an appearance on Bloomberg Tech, outlining her vision for how artificial intelligence should evolve with human oversight at its core.

World Labs, the AI startup founded by Fei-Fei Li, has acquired SceniX, a robotics company specialising in high-fidelity simulation systems. The move marks World Labs' biggest step into embodied AI, focused on enabling AI systems to perceive, reason, and act in physical spaces. SceniX builds simulation platforms where robots learn skills in virtual environments before deploying them on physical hardware. This "sim-to-real" approach fits with World Labs' Marble platform, which generates persistent 3D environments from text, images, or video. World Labs raised $230 million at launch in September 2024, followed by a $1 billion round in February 2026 that valued the company at $1.23 billion. Investors include Nvidia, AMD, and Autodesk. The company aims to develop spatial intelligence systems capable of understanding geometry, movement, and physics in physical environments.

Leading AI companies have secured billions in funding for "world models", a new frontier that aims to simulate the physical world rather than just process language. World Labs and Advanced Machine Intelligence each raised around $1 billion, whilst Runway secured $315 million. Unlike large language models (LLMs), world models create interactive 3D environments and simulations. They're being developed for robotics, scientific research, and asset generation for games and films. However, experts disagree on the precise definition. MIT's Vincent Sitzmann describes world models as systems that simulate future events given an interaction. World Labs' Ben Mildenhall emphasises real-time, continuous spatial understanding rather than LLMs' turn-based text exchanges. Most current approaches build on video generation technology, using autoregressive diffusion to create interactive, frame-by-frame simulations rather than pre-rendered sequences.