Decart AI was founded in 2023 in Tel Aviv by a team of researchers who had been working on the problem of real-time generative models — AI systems that don't just produce static outputs but generate interactive, continuous streams of content fast enough to feel like a live experience. The founding team, led by CEO Ido Shiraki, came from backgrounds in computer vision, GPU optimization, and neural network architecture, and they converged on a provocative question: what if you could run a world model fast enough that it replaced a traditional game engine entirely? Not as a pre-rendering tool or an asset generator, but as the runtime itself — generating every frame, every physics interaction, every visual response to player input in real time. That question became Decart's founding thesis and led to one of the most attention-grabbing demos in generative AI.
In late 2024, Decart released Oasis, an AI model that could generate a playable Minecraft-like experience in real-time, entirely through neural network inference. There was no traditional game engine, no pre-built world geometry, no physics simulation — just a transformer-based model generating every frame based on the player's inputs, running at interactive frame rates. The demo was immediately viral. It was rough around the edges — visual artifacts, inconsistent physics, limited world persistence — but the fundamental achievement was undeniable: a neural network was generating a coherent, interactive 3D world fast enough that you could walk around in it. The technical feat required extraordinary inference optimization, squeezing generation latency down to the roughly 50-millisecond budget needed for 20+ frames per second. Decart published the approach and open-sourced a version of the model, which only amplified the buzz.
Decart's work sits within the broader "world model" research direction that gained significant momentum in 2024-2025, championed by figures like Yann LeCun at Meta and explored by multiple labs including Google DeepMind, Runway, and World Labs. The core idea is that AI models should learn an internal representation of how the world works — physics, object permanence, cause and effect — rather than just pattern-matching on static data. What makes Decart's approach distinctive is the emphasis on real-time interactivity. Most world model research focuses on video generation or planning, producing outputs that you watch rather than interact with. Decart's models are designed to respond to continuous input, making them more like game engines than video generators. This interactive dimension is technically far more demanding but also far more commercially interesting for applications in gaming, simulation, training, and robotics.
Decart raised $21 million in seed funding in 2024, led by Sequoia Capital with participation from notable investors including Nvidia's venture arm. For a seed round, this was substantial, reflecting investor enthusiasm for the world model space and the viral impact of the Oasis demo. The company's immediate technical challenge is closing the gap between "impressive demo" and "production-quality experience" — the generated worlds need better consistency, longer coherence windows, and the kind of visual fidelity that players and users expect from modern game engines. The longer-term opportunity is much bigger than gaming: real-time world simulation has applications in autonomous vehicle training, robotic manipulation, architectural visualization, and any domain where you need to generate realistic interactive environments on the fly. If Decart can make their inference fast enough and their outputs reliable enough, they could define an entirely new category of AI-native interactive media.