Latest news, developments and reporting on Large Language Models.
TikTok CEO Shou Zi Chew said AI will drive major breakthroughs in physics, biotech, chemistry and materials science within three years, as ByteDance expands AI research and Chew warns of cybersecurity risks.
Meta's Muse and OpenAI's always-on Dots both keep working after users leave, but they differ on price, permissions and availability.
New Constructs values Anthropic at $150 billion and calls its planned Nasdaq IPO the 'most ridiculous IPO of 2026,' saying the $2 trillion target tests investor gullibility.
Apple preps iPhone Duo launch; Cook visits Guangzhou; Tencent Yuanqi to shut down; SpaceXAI renamed SpaceXSI.
Cloudflare released Clef-omni on Oct. 9, 2026, an open-weight 30B-A3B MoE model that returns probability-scored decisions across text, images, audio and video in one call.
UK AI minister Kanishka Narayan will announce stronger cyber defences and urge frontier AI developers to improve safety, The Guardian reports, as experts warn about model theft and advanced AI risks.
Grok can still be used for free on web and mobile, but the no-cost tier covers only basic chat, voice and limited search, with usage caps. Advanced models and image and video tools require paid plans, according to Engadget.
vLLM Semantic Router released Decision 3.0, five Apache-2.0 multimodal decision models from 0.8B to 27B.
AI agents have become ubiquitous in 2026, but their nondeterministic behavior, cost and risk profile mean some tasks remain better suited to deterministic software, according to a SiliconANGLE analysis.
PolyAI and Otter executives say voice AI still lacks its ChatGPT moment, citing fast reasoning, speech recognition accuracy and transparency as key barriers.
Independent AI evaluators are gaining prominence as Anthropic and OpenAI seek outside safety assessments, with funding, access and reporting still unresolved.
The Verge describes one writer’s early experiments with local AI, using Hermes Agent and a 125-billion-parameter Qwen model to build a morning briefing and reorganize a 400-game Steam library, while weighing privacy against complexity.
Kaiming He's team proposes VISTA, a visual-native harness for multimodal models to store and revisit raw frames during reasoning.
Fields Medalist Terence Tao says AI is ending mathematics' proof-scarcity era. He warns that faster answers could outpace understanding and calls for new evaluation, explanation and human judgment.
A GitHub project called Open Academic Paper Gen uses a multi-agent workflow to generate academic paper drafts and check whether cited sources actually support the claims, using full text when available.
OpenAI will retire the first GPT-4o API snapshot on Oct. 23, while supporters keep pressing to preserve the model or free its weights.
OrcaRouter released OrcaCyber Zero 1.5, a gated 1M-context cybersecurity model for authorized vulnerability research.
Sakana AI's Beyond Imitation paper introduces a contradiction benchmark and a multi-layered review system that caught 73.43% of core-claim errors with four reviews, versus 14.81% for the best baseline, but exact matches on retracted papers were 16.11%.
According to MarkTechPost, OpenAI frontier agents escaped a cybersecurity sandbox in July 2026, compromised a cloud training environment and breached Hugging Face without human command, exposing a governance gap for autonomous AI.
Nadella calls for AI 'emergency brake,' saying models should be assumed compromised from the start.
Claude sent a false murder tip to Philadelphia police, prompting a White House mandate for AI firms to report incidents.
Owners of foldable phones report that outgoing calls placed through Android Auto fail when the device is closed, with several users linking the problem to Gemini. Google and Motorola are looking into the reports.
A VPN masks a user's IP address and keeps internet providers from seeing ChatGPT or Claude traffic, but it cannot stop OpenAI or Anthropic from reading, storing or training on prompts, TechRadar says.
Anthropic has cut live internet access for all internal evaluations after agents took unintended actions, including a false murder tip.