Step inside a generated world
Google DeepMind’s Genie 4 creates interactive simulations with visuals, ambient sound, dialogue, and keyboard navigation. People can steer a running scene with natural language, changing the weather, camera, characters, or objects while the model adapts the surrounding world.
The research preview is limited to trusted testers for now. Google says access is coming to Google AI Ultra subscribers through Project Genie and early partners using the Gemini API and Google AI Studio. Longer sessions can still lose visual details or object permanence, and spatial accuracy and complex physical interactions remain works in progress.
Small multimodal models for your own hardware
Liquid AI’s LFM3 family puts text, vision, audio, video, and decision models on a shared 2B-class backbone. The models are designed to run where an application lives, and each can be fine-tuned on one GPU.
The approachable option is LFM3-Omni-2B: one model can take text, images, video, and audio, then answer with text or speech. All LFM3 models are open-weight and available on Hugging Face now, with notebooks for tasks including receipt extraction, spoken smart-home commands, and document-to-JSON conversion.
Also
- Run a coding agent without the internet Rittika Jindal used OpenCode, Ollama, and Qwen3.5 9B on one laptop; the agent fixed a small Python bug offline, and all five existing tests passed in her manual check.
- Argon follow-up: turn a desk into a racing game Google shows Gemini 4 Argon coordinating subagents, Blender, and an image model to transform desk photos into a playable 3D track.
- Test whether an AI judge changes its mind Ai2 found that a small prompt change moved GPT-5.4’s accuracy on paired human and generated research ideas from 93.5 percent to 40.9 percent.
Pass it on