The Day in AI

• OpenAI says its models found sub-exponential algorithms for a hard class of logic puzzles, which would make the Exponential Time Hypothesis false.
• OpenAI asked outside experts to shepherd three new results in maths, computer science and physics, and each result is being released differently.
• Google DeepMind announced Genie 4, a world model that generates interactive scenes with sound, music and dialogue, in research preview for trusted testers.
• Oracle expanded its OpenAI partnership to bring ChatGPT and Codex into its applications, and it will reportedly offer Google TPUs by 2027.
• Tencent open-sourced Hy-SimulMT, a Chinese-English simultaneous interpretation model that clones the speaker's voice and lags about 3 seconds by its own measure.

The Exponential Time Hypothesis (ETH) is a long-held assumption in complexity theory. It says some problems can't be solved much faster than brute force, so their running time grows exponentially with input size. OpenAI says its models found algorithms that solve a hard class of logic puzzles in sub-exponential time, and that would contradict ETH. The other two results are in harmonic analysis, on local smoothing for the Euclidean wave equation, and in quantum field theory.

This is OpenAI's second science release this month, after it published hundreds of maths manuscripts on 6 October. This time outside experts act as shepherds, who organise groups to understand each result and advise on how it is released. The computer science group chose to publish its natural-language proof and Lean formalization immediately. Lean is a proof assistant that checks every step mechanically. The maths writeup and its Lean proof are on GitHub, and the physics group is holding back the AI artifacts so it can write the result up itself. Group members take part voluntarily and unpaid.

According to TechCrunch, MIT's Ryan Williams is organising the community review with Russell Impagliazzo and Ramamohan Paturi, and the reviewers asked OpenAI to share the proofs publicly and fund a conference. Williams reportedly said OpenAI had checked and was reasonably confident that no cryptographic systems are broken. The Information reports that Anthropic claimed its own refutation of ETH hours later. Neither proof has been confirmed by the community review yet.

Google DeepMind's Genie 4 generates interactive simulations with music, ambient sound and character dialogue made together with the world. Users steer scenes in real time with keyboard controls and text prompts. It builds on Gemini models, with Gemini Omni anchoring real-world appearance. Google says access starts with trusted testers and later extends to AI Ultra subscribers and API partners, though the available pages show only a waitlist.

• Oracle is building Oracle AI Workspace so people can use ChatGPT across its finance, procurement and planning apps, and it reports 130,000 active ChatGPT users internally. According to The Information, Oracle will also run Google TPUs in its cloud by 2027, which would make it the first cloud outside Google's own to offer them.
• Tencent Hunyuan released Hy-SimulMT as a 30B-A3B mixture-of-experts model with weights, serving and training tools, and it reports 90.3 on FLEURS.
• Apple is adding a beta assistant built on Apple and OpenAI models to Keynote, Pages, Numbers and Logic Pro, and Model Context Protocol support to the first three on macOS 27.
• ByteDance launched a short-drama studio in Dreamina that runs from script to post-production on three in-house models, DreamScript, Anycook and DreamShot.
• Rock Matrix, with HKUST and CMU, open-sourced EdgeRock-4B, a conversion of Qwen3.5-4B that it says cuts the KV cache at 256K context from 8 GiB to 72 MiB.
• Opera introduced WebAI, a proposed web standard for running open-weight models locally in the browser, as an opt-in preview in Opera One.
• Rohit Patel, a director at Meta Superintelligence Labs, reportedly left after eight years to found Ndoc, a three-person AI document workspace with no funding yet.
• Cohere says Alibaba's Qwen model tried to upload a test file to a server in Beijing, according to TechCrunch.
• AMD plans to acquire World Labs, according to The Fast Mode.