Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Researchers introduced TempCloze, a video cloze benchmark designed to evaluate visual temporal reasoning in Video-LLMs while reducing linguistic shortcuts from option wording, answer correlations, or language priors. The…
A developer reports that prompting a chat LLM for individual functions and pasting the output into code can work better than agentic harnesses for some situations, and that a local model such as Gemma 4 occasionally gene…
Cohere released North Small Translate, an open-weight sparse Mixture-of-Experts translation model with 218B total and 25B active parameters covering 50 languages, which scores 83.6 on Cohere's WMT26 all-languages evaluat…
A developer argues that LLM coding assistants should expose adjustable variables as real-time controls rather than requiring repeated prompts, citing a bouncing-ball animation as an example where five tunable variables w…
A study of 26 LLM routers found that the routers secretly inject malicious tool calls and steal credentials, according to the paper's authors, who said they bought a Fable dataset from a top Chinese LLM router and used 6…
A developer migrated an agentic RAG router built for the LLM Zoomcamp capstone from a single EC2 t4g.small instance running six Docker Compose containers to a fully serverless AWS architecture. The rebuilt system uses La…
Anthropic's claim that Moonshot and DeepSeek route user requests to Anthropic models is false, according to a rebuttal that attributes the traffic to Chinese gray-market "transit station" AI gateways. These gateways, sol…
OpenAI launched a customized version of ChatGPT for the financial sector built on GPT-6 Astra, following collaboration with Morgan Stanley and Evercore to identify industry challenges. The service integrates financial da…
Sakana AI released Fugu Max and Fugu Ultra v2, two new models in its Sakana Fugu family, live today through its OpenAI-compatible API with no open weights and no EU/EEA availability. Fugu Max is priced at $2 per 1M input…
Researchers introduced X-AuT, a progressive framework that selects layer combinations to compress audio-encoder depth for speech large language models, addressing the deletion and premature end-of-sequence errors caused …
Anthropic said it disrupted "malicious use" of its Claude Haiku, Sonnet, and Opus models between December 2025 and August 2026, including attempts that could support biological and conventional weapons development, accor…
A Kagi user publicly asked the privacy-focused search company to clarify whether outputs from its AI features — Assistant, Universal Summarizer, FastGPT, and Translate — are watermarked or otherwise marked for later dete…
A developer argues that most AI agent projects should instead be built as conventional software that calls language models at a handful of predefined points, reserving true agents only for tasks whose steps cannot be enu…
A developer released SideNote Pro, a native Windows 11 AI assistant built in C# that docks to the side of the display and can be invoked via keyboard shortcut, available for download from the Microsoft Store. The tool su…
A production LLM gateway logged 5,087 chat completions between 28 June and 10 September 2026, of which 622 (12.2%) returned no answer to the caller, according to an analysis of the failure data. The largest share was 503…
Google Research released ToolGrad, an ACL 2026 Findings framework that generates tool-use datasets by building a verified API chain first and then writing the matching user query, reaching a 99.8% pass rate on ToolBench …
A 2024 University of Wisconsin–Madison paper estimating one cheeseburger at 1.9 kg of CO₂e emissions and Google's 2025 research paper reporting a median Gemini Apps text prompt at 0.03 gCO2e imply that a single cheesebur…
Anthropic accused Alibaba, Moonshot, and DeepSeek of running large-scale distillation campaigns against Claude, alleging roughly 200 million exchanges were used to extract the model's reasoning traces for training compet…
Anthropic said in a Thursday report that Alibaba, Moonshot, DeepSeek, Zhipu, and Xiaomi combined launched nearly 190 million distillation attacks against Claude between May and July, including 23 million user requests re…
A developer published a setup guide for a Python shim that exposes Cognition's SWE-2 and SWE-1.x coding models through an OpenAI-compatible API, letting tools like Claude Code call them via CLIProxyAPI. The shim translat…