Latest news, developments and reporting on Large Language Models.
Google has released Gemini 4 Argon, a flagship model it says beats top OpenAI and Anthropic offerings on some benchmarks, with a phased rollout starting with trusted cybersecurity partners and U.S. safety reviews.
Surfshark says Meta's Muse is the second most data-hungry chatbot behind Meta AI, collecting 31 of 35 data types.
MIT Technology Review essay argues LLMs do not truly reason, contrasting next-token prediction with AlphaGo's search and intuition.
openJiuwen has released X-Router, a self-evolving routing engine that decides which model handles each agent request. The team reports token consumption down more than 50% in tests and describes the release as Ascend-friendly.
A new paper co-authored by Shing-Tung Yau claims all 28 seven-dimensional spheres admit strictly positive sectional curvature, and thanks GPT 6 Astra and Claude Pro for help with part of the proof strategy and calculations.
AWS Strands Labs has released Strands Decider 2B, a 1.9-billion-parameter open source model that returns a choice, a yes/no probability or a score instead of text. Weights ship under Apache-2.0, and it runs locally with a median latency of 115 ms on an RTX 3090.
MiniMax's newly launched M3.1 Flash Preview produced front-end design, animation and game prototypes comparable to Claude Opus 5.5 when tested with the same public prompts, according to ifanr.
ServiceNow CoreAI's AutoSynthData generates validated training tasks from enterprise agent failures.
ChatGPT now offers virtual clothing try-on from a selfie or full-body photo, and adds Favorites and folders for saved products.
Microsoft introduced MAI-Transcribe-2-Streaming and two MAI-Voice-2.1 text-to-speech models for developers building near-instant voice agents, with streaming transcription priced at $0.54 per audio hour.
Cloudflare has released Clef and Clef-flash, the Workers AI team's first trained models: Apache 2.0 decision models that return probabilities for typed questions instead of generated text and are compatible with TypeSafe AI's Jev API.
Study: AI chatbot answers nearly 19% less diverse than Google Search, raising knowledge-narrowing concern.
Satlyt Inc., founded in 2024 by former SpaceX Starlink product manager Rama Afullo, has raised $8 million led by Non Sibi Ventures for AI software that processes satellite data in orbit.
Anthropic reportedly targets pre-Thanksgiving IPO at over $2 trillion valuation despite losses and AI-risk warnings.
OpenAI has dismissed three safety researchers accused of sharing confidential material with an outside AI safety group, the company confirmed. The firings follow earlier safety departures and a string of incidents involving OpenAI models.
Cloudflare has released Clef and Clef-flash, open-weight decision models for structured tasks, images, video and 64K contexts, claiming benchmark gains over TypeSafe's Jev.
Cognition and CoreWeave executives said always-on AI agents such as Devin require infrastructure that continuously links inference, feedback and training, as CoreWeave announced Forge to support that loop.
OpenAI launched ChatGPT virtual try-on and a Favorites library, both using its new ChatGPT Images 2.5 model.
Google launches Guided Vision in Gemini Live for Android, giving blind and low-vision users spoken camera descriptions.
Google this week unveiled Gemini 4 Argon, a frontier model it says advances coding, cybersecurity and complex tasks, and launched a prototype satellite for Project Suncatcher. Investors and consumers are pressing Google for a breakout personal agent, where Meta and OpenAI are moving fast.
Le Monde has published an article asking who Samuel Fitoussi is, describing him as the conservative columnist behind accusations that Orélien's best-selling novel was written with artificial intelligence.
OpenAI has dismissed three safety team researchers over allegedly sharing confidential information with an outside AI safety organization, The Wall Street Journal reports.
Graphite study finds AI models still have distinct writing tells, such as 'this matters' and corrective framing.
Google released Gemini 4 Argon, a flagship model it says leads several benchmarks, costs less on some tasks and is initially limited to vetted cybersecurity defenders.