Latest news, developments and reporting on Large Language Models.
The fourth AFAC Financial AI competition concluded in Shanghai with over 5,000 teams and 20,000 contestants, and organizers released a huge financial AI dataset.
World Labs released Atlas on Sept. 2, but Chinese startup Yingsu had already open-sourced a comparable world model in March and later ranked first in WorldArena 2.0. Yingsu also launched an open data initiative.
AI-generated food images are unsettling diners, and Instagram's AI Content tags are appearing on ordinary photos. Technical limits in diffusion models and vague detection standards are driving the confusion.
Nvidia agrees to acquire AI hub Hugging Face for $12.93B, vowing to keep the platform open and neutral.
Four major AI models—ChatGPT, Claude, Grok, and Gemini—suffered overlapping outages on Thursday morning, a rare simultaneous failure; OpenAI and Anthropic recovered while Grok was still impaired.
Nvidia confirmed its $12.93B acquisition of Hugging Face, gaining an open-source AI hub and model trend visibility.
Four US agencies sued over secret AI testing framework; watchdog seeks disclosure by Sept. 30.
A TechRadar opinion piece by Lance Ulanoff says GPT-6 is here and asks whether we should decline Astra, a model described as too powerful to understand.
Meta launched Muse Spark 1.3, an agentic coding model with 20% fewer tool calls and 25% fewer tokens.
Meta offers Muse Spark users roughly 95% off if they share prompts and outputs to train future models, TechCrunch reports.
OpenAI has released GPT-6 Astra, showing near-perfect scores in math and cybersecurity, while its leaders declare the AGI era has begun.
Google brought Live voice AI to Gmail, Docs and Keep, and released Gemini 3.8 Flash, calling it its best reasoning and coding model yet.
Google has introduced WeatherNext 3, an AI weather model that taps live satellite data for hourly, higher-resolution forecasts with better precipitation accuracy.
OpenAI's GPT-6 is set to adopt recurrent-depth reasoning, hiding the chain-of-thought, while Russian startup Mostik demonstrates direct model-to-model communication without generating tokens. These advances undercut the token-based AI pricing model.
Grok and other AI services went down Thursday after a SpaceXAI data center outage; SpaceXAI apologized and says systems are restored.
Anthropic confirmed Sept. 3 outages on Claude models including Opus 5 and Mythos 5.1; a fix is underway.
Hugging Face introduced NeoMME, a 260M/800M-parameter multimodal encoder using a single bidirectional Transformer over text and images, enabling efficient visual document retrieval with compact indexes.
Leiphone reports two arXiv preprints: Gemini 2.0 proves the strongest VLM judge in embodied tasks, enabling the low-cost VLM-AR3L framework, while a 2B-parameter S²-VLA beats 7B models in long-horizon robot manipulation using adaptive attention.
A Nanjing University associate professor sparked heated debate by saying computer science students without paid AI tokens should drop out — and requiring his students to self-fund tokens and avoid “ancient programming” in class.
Anthropic accuses Chinese AI labs of illegally distilling its Claude models, a practice its security chief calls a national-security risk.
A Sun Yat-sen team's SpatialSV embeds explicit 3D geometry in MLLMs, improving spatial accuracy and halving manual annotation.
ChatGPT, Claude and Grok went down at nearly the same time on Sept. 3, forcing companies to apply fixes.
Anthropic's Claude 5.1 separates model capability from execution permissions in new Fable and Mythos tiers, Leiphone says.
O'Reilly's chief product officer argues that AI-era free content makes publisher credibility more valuable, not less.