Latest news, developments and reporting on Large Language Models.
SpaceXAI is joining the Omacom Foundation as a founding corporate patron and donating $1.5 million in Grok tokens to David Heinemeier Hansson's Omarchy Linux project, The Verge reported.
Google unveiled a unified Gemini agent that can plan and complete work on a user's behalf, launching first for businesses before a later consumer rollout, per a TechCrunch report.
Anthropic updated its usage policy on Thursday, banning sustained abuse of Claude and codifying restrictions on election interference, weapons software and surveillance.
JetBrains has released Mellum2.1, a 12B open mixture-of-experts reasoning model with 2.5B active parameters for coding agents and self-hosting. It beats Mellum2 on 15 of 17 benchmarks but still trails Qwen3.5-9B on several hard agentic and knowledge tasks.
Claude Haiku 5.5 lifts OSWorld 2.1 to 72.4%, cuts average cost about 75%, adds adaptive reasoning and a 1M-token context.
TechRadar reports OpenAI has quietly added an Intelligent UI to ChatGPT, powered by GPT-6, that turns ordinary prompts into mixed text, visual and interactive answers. The feature is available in the standard Chat tab to all users, including those on the free tier.
Meta and OpenAI have launched consumer AI agents, Muse and Dots, as always-on assistants that can book reservations, triage email and take on workplace tasks. The Verge reports the race raises questions about trust, privacy and Meta's lack of a frontier model.
Google has released Google AI Edge Foresight, a free experimental Mac app that transcribes meetings entirely offline using an on-device EmbeddingGemma 2 model. It competes with Granola and other AI note-takers.
Google Cloud launched Gemini agent, an autonomous AI assistant that works across apps and devices.
MIT Technology Review’s Oct. 8 Download says humanoid robot hype is outrunning AI’s grasp of physical variability, while WaveSave’s portable rubber dam is named a climate tech company to watch.
Google Maps has launched its first fan-favorite dining list, using reviews, ratings and directions to spotlight popular restaurants and food trends in 10 cities, while adding Gemini-powered discovery and ordering through Toast, Square and Uber Eats.
Leiphone published a seven-layer cross-domain AI transfer governance codex, while QbitAI reported that UniPat AI's PaperBenchX benchmark found GPT-6 Astra fully reproduced only 13.98% of 93 research papers.
Hu Han, formerly of Tencent Hunyuan, has a new multimodal startup that plans to raise tens of millions of dollars at a target valuation of hundreds of millions, Leifeng reports. Yuanshi Capital is the financial adviser.
NVIDIA researchers, with Princeton University and the University of Maryland, introduced PivotOPD, an on-policy distillation method that trains multi-turn LLM agents to avoid pivotal mistakes and recover from them. It beat 13 baselines on ALFWorld, WebShop and Search-based QA.
Anthropic data shows a state-sponsored campaign scored maximum risk while looking ordinary by technique count.
Shengshu Technology has opened Vidu Q4 Preview, a high-expression video-generation model priced from 0.09 yuan per second, with support for up to 15 reference images and 2K or 4K output, according to ifanr.
Alibaba said at its annual Yunqi Conference that future Qwen models will scale toward 5 trillion to 10 trillion parameters, while new multimodal and AI-for-chip-design efforts aim to speed development and lower costs. Leiphone reported.
OpenAI is rolling out GPT-6 to every ChatGPT tier with an Intelligent UI that builds interactive replies; paid users get GPT-6 Sol, Free and Go accounts get the less capable Luna.
OpenAI says an unreleased internal model has produced 372 results resolving or substantially advancing open problems in mathematics and theoretical computer science, with many verified in Lean. The company disclosed the results in a GitHub repository, and mathematicians are still parsing them.
Perplexity released pplx-embed-v2-late, ColBERT-style multimodal embedding models in 0.6B and 9B sizes, with MIT-licensed weights on Hugging Face. The 9B model scored 92.4% on MADQA, while the 0.6B model is designed for edge query encoding.
Huawei and Nvidia target KV cache storage; Ruqi builds real-world data infrastructure for physical AI, Leiphone reported.
CAIR released CARES 4.0, a multimodal clinical agent system, at the fourth CREATE Symposium in Hong Kong, where clinicians, hospital leaders, researchers and investors discussed medical AI's path into clinical workflows and commercialization.
US venture deal value reached a record $515.8 billion in the first nine months of 2026, lifted by OpenAI and Anthropic mega-rounds, while exits lagged far behind, PitchBook-NVCA data show.
Jieyue Terminal will launch STEPX Neo, its first large-model-native agent phone, in Shanghai on Oct 13.