Releases·Model Release
Runway's GWM Worlds 2 Adds WorldPrompt for Real-Time Interactive World Generation
Runway introduced GWM Worlds 2, a research preview that turns high-fidelity video and audio generation into real-time interactive simulation, and detailed its new WorldPrompt input format in interviews with CTO Kamil Sindi, Principal Research Scientist Robin Kahlow, and co-CEO Anastasis Germanidis. WorldPrompt lets users fix parts of a simulated environment and prompt timestamped, real-time events, differentiating Runway from Genie 3, Odyssey-2 Pro, and World Labs' RTFM. The work matters to practitioners because real-time autoregressive world generation remains an open engineering problem with limits in error accumulation, memory, and counterfactual realism.
0 engagement1 sourceFri, Sep 25, 2026, 01:30 AM
AI Overview
- Runway introduced GWM Worlds 2, a research preview generating real-time interactive 720p video at 24 fps and 48,000 Hz audio via autoregressive diffusion
- WorldPrompt lets users fix environment aspects like the first frame and add timestamped, real-time promptable events controlling characters, cameras, and environment
- CTO Kamil Sindi and scientist Robin Kahlow said movement prompts work reliably but the preview has flaws and lacks perfect long-term memory
- Co-CEO Anastasis Germanidis said distillation cut denoising steps from about 50 to four, calling error accumulation the biggest autoregressive challenge
- Runway, valued at $5.3 billion after a $315 million February raise, competes with Google DeepMind's Genie 3, Odyssey-2 Pro, and World Labs' RTFM
- Kahlow cited robotics testing and running thousands of simulated agent environments, with no structured state—agents observe only video and audio
- Germanidis said counterfactual generation, like equally realistic failed and successful goals, marks the gap between video models and world models
Related
Releases·Model Release
Runway launches GWM Worlds 2 research preview for real-time interactive world simulation
Runway announced GWM Worlds 2, a research preview for real-time interactive world simulation built on its foundational audio-visual generation model. The system generates continuous 720p video at 24 fps with 48 kHz audio that responds dynamically to user inputs like movement, dialogue, and weather changes, rather than relying on pre-scripted clips. It matters to practitioners as an early playable world model in the growing real-time generative simulation space.
Runway announced GWM Worlds 2, a research preview for real-time interactive world simulation built on its foundational audio-visual generation model. The system generates continuous 720p video at 24 fps with 48 kHz audio that responds dynamically to user inputs like movement, dialogue, and weather changes, rather than relying on pre-scripted clips. It matters to practitioners as an early playable world model in the growing real-time generative simulation space.
69 engagement1 sourceredditThu, Sep 3, 2026, 06:01 PM
Releases·Model Release
AlayaWorld open-source video world model released with 720p 24 FPS streaming and camera control
AlayaWorld is a full-stack, open-source video world model that supports 720p, 24 FPS streaming video generation with camera control and text-driven event generation. It is built around four core properties: interaction, consistency, stability, and runtime.
AlayaWorld is a full-stack, open-source video world model that supports 720p, 24 FPS streaming video generation with camera control and text-driven event generation. It is built around four core properties: interaction, consistency, stability, and runtime.
64 engagement1 sourceredditWed, Jul 22, 2026, 03:46 AM
Releases·Model Releasefeatured
World Labs releases Atlas, a spatial world model generating unseen viewpoints and navigable space-time simulations
World Labs announced Atlas, a foundation model for spatial intelligence that ingests text, images, image sequences, camera poses, and depth to build persistent spatial understanding of scenes. Built on an autoregressive diffusion transformer, it generates unseen viewpoints and infers missing geometry rather than only reconstructing observed content, and converts scenes into Gaussian splats for fast rendering. From footage captured on just 3-5 consumer phones it can build complete navigable space-time simulations, positioning it as a step toward physical AI.
World Labs announced Atlas, a foundation model for spatial intelligence that ingests text, images, image sequences, camera poses, and depth to build persistent spatial understanding of scenes. Built on an autoregressive diffusion transformer, it generates unseen viewpoints and infers missing geometry rather than only reconstructing observed content, and converts scenes into Gaussian splats for fast rendering. From footage captured on just 3-5 consumer phones it can build complete navigable space-time simulations, positioning it as a step toward physical AI.
210386 engagement3 sourcesredditTue, Sep 1, 2026, 07:25 PM
Releases·Model Release
Astronex-World 1.0: Real-Time Interactive World Model Foundation
Astronex-World 1.0 is an open controllable video world-model foundation supporting text-to-video and image-to-video generation under camera trajectories, continuous actions, and inserted text events. Built on a 5B prior, it streams 832x480 video at 24 fps in real time on a single GPU and outperforms larger 13.6B–14B models on WBench. Its reserved action interfaces enable post-training for embodied intelligence and autonomous driving.
Astronex-World 1.0 is an open controllable video world-model foundation supporting text-to-video and image-to-video generation under camera trajectories, continuous actions, and inserted text events. Built on a 5B prior, it streams 832x480 video at 24 fps in real time on a single GPU and outperforms larger 13.6B–14B models on WBench. Its reserved action interfaces enable post-training for embodied intelligence and autonomous driving.
0 engagement1 sourcearxivThu, Sep 17, 2026, 10:38 AM
Social·Community
World Models Need Causality, Not Pretty Pixels — Christopher Manning, Moonlake AI
Stanford professor Christopher Manning, now at Moonlake AI, gives a talk tracing AI history from Dartmouth 1956 to LLMs and arguing that embodied AGI should be built on simulation. He contends generative video world models like Genie 3 simulate observations without underlying semantics and thus can't support planning, and describes Moonlake's approach: building action-conditioned, code-based world models from a single photo or short video.
Stanford professor Christopher Manning, now at Moonlake AI, gives a talk tracing AI history from Dartmouth 1956 to LLMs and arguing that embodied AGI should be built on simulation. He contends generative video world models like Genie 3 simulate observations without underlying semantics and thus can't support planning, and describes Moonlake's approach: building action-conditioned, code-based world models from a single photo or short video.
19 engagement1 sourceyoutubeThu, Sep 24, 2026, 01:30 PM