AI News Briefing — October 28: OpenAI and Anthropic models disprove the Exponential Time Hypothesis, Google unveils Genie 4

Today's Headlines

• OpenAI's models produce three proofs in math, computer science and physics; Anthropic also publishes a refutation of the Exponential Time Hypothesis
• Google DeepMind unveils Genie 4, a world model that generates sound and dialogue along with the scene
• OpenAI and Oracle expand their partnership to bring Oracle's apps and data into ChatGPT

The day belonged to proofs written by AI. OpenAI announced results from its internal models in three fields, and for one of them, the claim that the Exponential Time Hypothesis is false, Anthropic published its own proof reaching the same conclusion.

Google DeepMind introduced Genie 4, a world model that builds an interactive world in real time as the user steers it.

Today's Top Three

OpenAI's models produce proofs in math, computer science and physics, and Anthropic also refutes the Exponential Time Hypothesis

OpenAI said on October 26 that its internal models have produced new proofs in mathematics, computer science and physics.

The computer science result shows that the Exponential Time Hypothesis (ETH) is false. The ETH holds that deciding whether a Boolean formula can be satisfied takes time exponential in the number of variables. It has been widely believed and used as a premise for conditional hardness results in hundreds of papers, across graph theory, computational biology, databases and other fields. OpenAI says its model found the disproof on October 16 and formalized it in the Lean proof assistant the next day.

A group led by Russell Impagliazzo and Ramamohan Paturi of UC San Diego, who proposed the ETH, and Ryan Williams of MIT will verify and explain the result. At the group's request, OpenAI published the natural-language proof and the Lean proof on GitHub and will fund a conference on the implications. Participation is voluntary and unpaid, and OpenAI says it is not seeking academic credit for the discoveries.

In mathematics, OpenAI says its models proved Sogge's 1991 local smoothing conjecture, which concerns how waves focus, in every dimension, which also settles the Kakeya conjecture and Stein's restriction conjecture in all dimensions. The AI output and the Lean proof were released as they are.

In physics, OpenAI says its models showed that interacting quantum field theories in four dimensions, including quantum chromodynamics (QCD), exist in a mathematically rigorous sense. The work includes an existence proof for pure Yang–Mills theory, part of the Clay Mathematics Institute's Millennium Prize problem on quantum Yang–Mills theory. A group including David Gross, who shared the 2004 Nobel Prize in Physics, will steward the result; it has chosen to hold back the AI-produced materials and write up the work itself.

Anthropic also published a paper and a Lean formalization on GitHub stating that an internal research Claude model found a deterministic algorithm that solves 3-SAT in 2^o(n) time, which refutes the ETH. Anthropic describes the result as purely asymptotic, about how running time grows as the number of variables increases. Williams told TechCrunch that he asked OpenAI whether the result affects cryptographic systems, and that OpenAI had checked and was reasonably confident it implies no breaks.

• Working with academic communities to build scientific understanding (OpenAI)
• A 2^o(n) Time Algorithm for 3-SAT (Anthropic, GitHub)
• Frontier AI labs say their models refuted the Exponential Time Hypothesis. What does that mean? (TechCrunch)

Google DeepMind unveils Genie 4, a world model that generates sound and dialogue with the scene

Google DeepMind announced Genie 4 on October 27, a world model that generates an interactive world in real time in response to the user's input.

Genie 4 generates music, ambient sound and the speech of characters together with the video. Users move and steer the camera with a keyboard, and can type instructions such as changing the weather or bringing in a new person or object, which the world picks up as it runs.

Gemini models anchor the world's appearance and behavior. While Genie 4 renders, Gemini 3.8 Flash watches the scene, breaks goals into steps and checks that each step succeeded. For robotics research, DeepMind paired Genie 4 with Gemini 4 Argon to simulate Apptronik's Apollo humanoid playing a game of checkers.

Genie 4 is available first as a research preview for trusted testers, and is coming to Google AI Ultra subscribers in Project Genie and to early-access partners of the Gemini API and Google AI Studio. All video and audio it generates carries SynthID, Google's digital watermark.

• Genie 4: A new generation of world simulation (Google DeepMind)

OpenAI and Oracle expand their partnership to bring Oracle's apps and data into ChatGPT

OpenAI said on October 27 that it is expanding its partnership with Oracle so that Oracle applications and data can be used inside ChatGPT.

Oracle is building Oracle AI Workspace, which connects ChatGPT to Oracle's business applications. The aim is to let employees pull together information from finance, procurement and planning apps and complete approved workflows within their organization's permissions and approvals.

Oracle also announced products built on OpenAI models. One of them, Oracle AI Security Mesh, uses OpenAI's cyber models to find and fix vulnerabilities in applications and infrastructure, with Oracle managing the testing environment.

Oracle customers in the U.S. can buy access to the OpenAI API, Codex and ChatGPT Work through Oracle Marketplace, using existing procurement processes and eligible Oracle Universal Credits.

According to OpenAI, Oracle has 130,000 active ChatGPT users and more than 95,000 active Codex users. Its recruiting team now prepares market research in 15 to 20 minutes that once took two to four days.

• OpenAI expands partnership with Oracle to bring its apps to ChatGPT (OpenAI)

Other Developments

Models

• Liquid AI released LFM3 on October 27, an open-weight family of six models of about 2 billion parameters each, covering text, vision, audio and more. All six branch from the same LFM3-1.7B-Base backbone, support a 256K-token context and are built to run on devices. Introducing LFM3: Every Modality for Specialized Intelligence (Liquid AI)

Products

• Apple added an AI assistant to Pages, Keynote and Numbers, rolling out in beta on devices that support Apple Intelligence. Built on models from Apple and OpenAI, it reads a document's context to draft text and build tables and charts, and on macOS 27 or later the apps accept outside AI agents through the Model Context Protocol (MCP). Apple is adding an AI chatbot to Pages, Keynote, and Numbers (The Verge)
• Opera introduced WebAI, a proposed web standard for running AI models inside the browser rather than in the cloud. It works with open-weight models such as Google's Gemma and arrives first as an opt-in early preview in Opera One. Opera introduces a new web standard to let users deploy AI locally in their browser (TechCrunch)
• Mozilla added two features to Smart Window, Firefox's AI browsing mode: filling web forms from the information in open tabs, and watching chosen pages for updates. Both are rolling out to the desktop beta in the U.S. and Canada. Firefox's Smart Window adds AI form filling and webpage monitoring (The Verge)

Research

• An Axiom Math project published a complete Lean formalization of the first-generation proof of the classification of finite simple groups, with every logical step checked by Lean's kernel. Along the way it found counterexamples to an assertion used in a published argument, and it has released a paper with an alternative argument. The Classification of Finite Simple Groups, formalized in Lean (Axiom Math)
• Researchers at TU Berlin and other institutions tested AI-personalized phishing on 7,741 university staff, building each message from information found online. Among the 3,310 people who received personalized emails, the click rate was 10%, about 2.7 times the 3.9% for a single generic message sent to everyone. AI-personalized phishing study, reported by ITmedia NEWS (Japanese)

Business

• Dallas-based investment firm Disruptive announced a $10 billion fund to back at least 10 late-stage companies in capital-intensive fields such as AI and defense, with checks of up to $1 billion. Disruptive raises $10B fund to make big bets on at least 10 companies (TechCrunch)
• General Compute, which builds its cloud on inference chips from Nvidia rivals, raised a $100 million Series A led by Blackbird Ventures and Abstract Ventures, at a post-money valuation of $800 million. General Compute raises $100 million Series A, countering the neocloud disaster narrative (TechCrunch)
• Rohit Patel, a former director at Meta Superintelligence Labs, left Meta after eight years and launched Ndoc, an AI-native workspace pitched as a replacement for Microsoft Office and Google Workspace. It places presentations, the reasoning behind the numbers, and the underlying data and Python scripts side by side so that AI-built figures can be checked. Meta AI director Rohit Patel leaves to take on Microsoft Office and Google Workspace (TechCrunch)
• Cybozu, a Japanese business-software maker, surveyed 551 companies and organizations with 1,000 or more employees: about 70% plan to increase AI spending over the next three years, while 29% track AI costs centrally across the company. Cybozu survey on AI cost management, reported by ITmedia Enterprise (Japanese)

Policy

• Candidates' positions on data center restrictions and moratoriums have become an issue in close U.S. midterm races in Texas, Pennsylvania and elsewhere, The Verge reports. Data center backlash is uniting America (The Verge)

Source: Selected by the editors from the AI news inbox (collected October 28, 2026: 55 items, 8 primary and 47 secondary).