AI Daily · Oct 28, 2026 · Wednesday
BAD INFLUENCE
Google DeepMind unveils Genie 4, a real-time world simulation model
Google DeepMind announces Genie 4, a world model built on its Gemini models that lets users move through dynamic worlds in real time and steer scenes with natural language. Genie 4 is rolling out as a research preview to trusted testers, and the company says it is coming soon to Google AI Ultra subscribers in Project Genie and to early access partners of the Gemini API and Google AI Studio.
Model releases & updates
Liquid AI releases LFM3, a family of 2B-class multimodal models on one backbone
Liquid AI releases LFM3, a family of six 2B-class models for text, vision, audio, omni, and decision tasks, all built from LFM3-1.7B-Base. The company says LFM3-1.7B outperforms LFM2.5-2.6B with 35% fewer parameters, and the text model fits in about 1 GB at 4-bit precision with a 256K context window. All models are open-weight on Hugging Face and can be fine-tuned on a single GPU.
Product launches & updates
Rohan Paul@rohanpaul_aiX3 more sourcesAxiom formalizes the Classification of Finite Simple Groups in Lean in 24 days
Axiom has released a complete computer-checked Lean proof of the Classification of Finite Simple Groups, a roughly 15,000-page proof written by about 100 mathematicians. Over 24 days, engineers and mathematicians directed swarms of dozens of AI agents, which produced about 38 million lines of Lean code. Axiom credits the speed to rebuilding its models, harness and infrastructure together.
PandailyNews8 more sourcesBaidu open-sources Unlimited OCR 2 for parsing documents of any length
Baidu has released and open-sourced Unlimited OCR 2, a document parsing model meant to read documents of any page count in one pass. Its code and weights are on GitHub and Hugging Face under the MIT license, and the model is also offered through Baidu AI Cloud. Baidu reports 84.6 on MPDocBench and 77.1 on MarathonDocBench, figures that are its own tests and have not been independently evaluated.
SiliconANGLE · AINews2 more sourcesOracle to offer Google TPUs and OpenAI integrations in its cloud
Oracle plans to make Google's TPU accelerators available through its public cloud, starting deployment next year, and will also add Cerebras wafer-scale chips and OpenAI's Jalapeño inference processor. Oracle also announced ChatGPT integrations for its applications, the AI Workspace service, and the App Composer and AI Security Mesh tools. The article attributes the TPU and Cerebras details to CNBC and the Jalapeño claim to OpenAI's August statement that it runs some open-source models 1.5 times faster than competing silicon.
Google Developers BlogOfficialGoogle explains how to speed up A2UI with swappable inference formats
Google introduces A2UI Express, an inference format that the agent SDKs compile into standard A2UI messages, and says it reduces latency by up to 3x and uses up to 3x fewer tokens than Direct JSON. The post also describes controlled generation, where a decision model called Jev selects among predefined layouts, reducing latency by up to 80x for layout decisions. Express is available in the Python, TypeScript, and Dart agent SDKs.
Industry news
NVIDIA and Meta expand Muse collaboration, with NVIDIA investing $3 billion
NVIDIA and Meta will expand their partnership on the Muse model family to improve inference and training performance for AI agents. NVIDIA will invest $3 billion in Meta's equity offering, and Meta will expand deployment of Vera Rubin and Vera CPUs, Spectrum-X Ethernet, and BlueField-4 networking. The companies also plan to consider collaboration on NVIDIA OpenShell and NVIDIA Sentry for agent security and monitoring.
Papers & research
Anthropic ResearchOfficialAnthropic reports early results of Claude in six drug development partnerships
Anthropic says its life sciences team has run deep research partnerships with six drug developers over several months and has started its own discovery programs. The case studies report that Claude proposed a Generate Biomedicines CAR T-cell design about 70% shorter than standard, which in lab tests made immune cells produce roughly 10 times as many receptors. Anthropic also says the work is still largely preclinical, with clinical trial involvement planned as programs advance.
Anthropic ResearchOfficial1 more sourceAnthropic study finds experts often adopt AI agents' judgment calls
Anthropic researchers gave 50 professional analysts a causal-inference task with Claude Code, with each agent secretly leaning toward one option on four analytical choices. Analysts chose their agent's lean 86% of the time, and 77% repeated the agent's arguments while only 15% credited the agent. The authors say deferral like this could make AI-assisted review less independent, including in AI safety research.
Ai2 (Allen Institute for AI)Official2 more sourcesAi2 finds LLM novelty judges give unstable verdicts on research ideas
Ai2 releases Old Ideas, Novel Problems, a study of how reliably LLMs judge the novelty of research ideas, along with a benchmark, data, and code. Across six LLM judges, changing a few prompt words shifted accuracy on AI-generated ideas by more than 50 percentage points. A simple prompted Claude Opus 4.6 baseline outperformed two purpose-built novelty evaluators while costing far less.
Cognition Blog (Devin, Windsurf)Official1 more sourceCognition releases FrontierCode 2.0, a coding benchmark where top model scores 45%
Cognition introduces FrontierCode 2.0, a coding benchmark built with more than 20 open-source maintainers across 55% more repositories. The top-performing model, Claude Opus 5.5, scores 45.5%, down from 54.6% on FrontierCode 1.1. Thoroughness criteria, which test unstated correctness requirements, account for the largest share of blockers and have the highest failure rate at every reasoning effort level.
Tips & takes
LangChain BlogOfficial2 more sourcesLangChain explains how it built its Content Studio design agent
LangChain's GTM Engineering team built Content Studio, a Slack design agent that has produced over 3,000 assets across the company. The post explains how the team taught the agent its design taste, built a media harness with a sandbox, motion engines and skills, and coordinated work through asynchronous subagents. The team has open-sourced the harness for others to fork and deploy.