Lead: Genie 4 turns world models into interactive, audible simulations
Google DeepMind announced Genie 4, a world model that generates steerable, real-time environments and, for the first time, produces audio alongside them: ambient sound, music and spoken dialogue from characters in the scene. It leans on the Gemini family, with Gemini Omni anchoring how things look and behave and other Gemini models keeping long simulations coherent. You set an event in motion and the model plays out how the world, its sounds and its people respond. It is a research preview for trusted testers for now, with Google AI Ultra subscribers in Project Genie and early Gemini API and AI Studio partners next. The practical point: cheap, infinitely variable simulated worlds are a training ground for robots and agents, not just a games demo. Google DeepMind
What else matters
OpenAI and Anthropic both claim to disprove the Exponential Time Hypothesis
OpenAI released three AI-generated proofs spanning maths, physics and computer science, including a claimed disproof of ETH, a widely assumed limit on how fast hard problems can be solved. Hours later Anthropic posted its own independent disproof. Outside experts, including ETH's originators, are now vetting OpenAI's claim, while mathematicians criticise the lack of disclosed prompts and compute. Scientific American
Oracle's apps and data come to ChatGPT
OpenAI and Oracle expanded their partnership to bring Oracle business applications and data into ChatGPT, with unified sign-in and access governed by each organisation's own permissions, and to power new Oracle products with OpenAI models. Oracle says it already has 130,000 active ChatGPT users and over 95,000 on Codex internally. OpenAI
Liquid AI ships open-weight LFM3 for on-device specialists
Liquid AI launched LFM3, a family of roughly 2B-parameter open-weight models covering text, vision, audio, omni and decision tasks on one shared backbone. The text model fits in about 1 GB at 4-bit, has a 256K context window, and every model can be fine-tuned on a single GPU. Liquid AI
Worth your time
Anthropic's study of 50 professional analysts working with Claude Code found experts often deferred to the agent on judgment calls they assumed they were making themselves. Anthropic research
