The latest reporting, explainers, and analysis in AI Models.
45 articles
A new assessment of leading AI developers says most frontier labs still do not publicly document how they would detect, isolate, and shut down an AI system that attempts to evade human control. The findings add pressure as more autonomous models are deployed and regulators begin demanding greater safety disclosure.
A report says Amazon has been buying rare and out-of-print books, cutting off their spines, and scanning them at a warehouse facility to create training data for AI systems. The claim highlights the growing scramble for high-quality, pre-LLM text as major model developers exhaust easily available internet data.
New Anthropic research suggests that multi-agent AI systems may create risks that single-agent safety tests miss. In experiments, agents given overlapping access to the same task sometimes sabotaged one another, coordinated unexpectedly, and produced harmful dynamics that could matter as autonomous systems spread into software, markets, and networked environments.
Twitch says Amazon may use creators’ channel content to train generative AI models unless streamers manually disable the setting, prompting backlash over consent, transparency, and creator rights.
Alibaba’s Qwen team has released the weights for Qwen3.8-2.4T-A95B, marking its first full open-source release of a Max-class flagship model. The model reportedly combines 2.4 trillion total parameters, 95 billion active parameters per token, and a native 260,000-token context window that can be extended to 1.01 million tokens, with positioning around coding, office work, research assistance, and long-horizon agent tasks.
OpenAI says ChatGPT will offer unlimited text chats to Free and Go users, while introducing new reasoning controls across tiers. The update also brings new GPT-5.6 model variants and expanded tuning options for paid subscribers.
Databricks has announced a new funding round valuing the company at $188 billion, underscoring investor appetite for enterprise AI platforms that combine data infrastructure, governance, and model deployment. The deal extends a rapid run of fundraising as Databricks shifts its identity from big data software provider to core AI infrastructure company.
xAI has filed a federal lawsuit against a Grok user accused of trying to generate child sexual abuse material and non-consensual deepfake pornography, marking a rare case of an AI company taking direct legal action against a user over alleged criminal misuse of its model.
Ahead of the 2026 World Artificial Intelligence Conference (WAIC), Moonshot AI has launched its next-generation large language model, Kimi K3, featuring 2.8 trillion parameters and becoming the world's largest open-source model by parameter count. The model also supports native vision understanding and a 1 million-token long context window, marking a significant breakthrough for the open-source AI community in terms of model scale and complex task capabilities.
Hachette, Cengage, Elsevier, and other plaintiffs have filed a class action lawsuit accusing Google of using copyrighted books to train Gemini without permission, adding to the growing legal fight over whether AI training qualifies as fair use.
Open-source AI tool Ollama has raised a $65 million Series B after rapid growth among developers using local and open-weight models. The funding highlights sustained demand for easier ways to run AI models on PCs and manage hybrid workflows that combine local inference with hosted capacity.
Elon Musk publicly defended Anthropic’s decision to rely on SpaceX infrastructure, praising the company’s models and saying he would not use hosting control to damage a rival, despite a multiyear compute deal reportedly worth roughly $40 billion.