Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Salesforce unveiled Koa, its first AI reasoning model, at Dreamforce on September 16, 2026, built on Nvidia's open-weight Nemotron family and post-trained for sales, marketing, and customer support tasks. Salesforce also…
A benchmark of llama.cpp v0.4.1-dev (commit 83078fe) on an Intel Panther Lake laptop found the OpenVINO backend unable to run the Qwen3.6-35B-A3B-MTP-UD-Q4_K_M.gguf model, failing with a quantized-V-cache flash-attn erro…
A new data page tracks release dates and training cutoffs for 20 current AI models across 8 labs, finding that only 9 of the 20 models have a training cutoff their lab actually publishes. The page lists models including …
Four open-source projects — Gilbeot, Sentinel, AirBridge, and Project Rosie — independently converged on the same pattern for making AI features reliable: rather than trusting the prompt, each changed what the model retu…
A developer built a minimal Python web server that turns Cerebras-hosted inference of Alibaba's Qwen 3.8 27B model into a live operating system with zero apps on disk, streaming tokens at 1,950 tokens per second. In the …
Mozilla and Mistral announced a partnership making Mistral models the engine behind Firefox Smart Window, Mozilla's beta AI browsing assistant, initially for users in France and North America with the United Kingdom and …
Pinecone, a fully managed vector database launched as a public beta in January 2021, is most commonly used today for retrieval-augmented generation (RAG) and agent memory, according to the company's positioning as an AI …
A developer measured token usage across 10 web pages and found that raw HTML consumed 3 to 24 times more tokens than the same pages converted to Markdown, with one product page arriving as 267,361 tokens of which only 3,…
Diogo Almeida, who helped develop ChatGPT and instruction-following research at OpenAI, launched Jev through his company TypeSafe, a System One AI model that does not generate text and instead returns typed probabilistic…
A developer argues that orchestrator skills for coding agents hit hard limits that no amount of better skill-writing can fix, because the procedure runs inside the agent session itself. The writeup identifies six such li…
A developer compiled a curated list of more than 35 text-generation LLM API providers that offer free, self-replenishing quotas, cataloging rate limits and daily token allowances for 47 services including Google Gemini, …
A developer using Claude Code on a Max subscription documented how making the top model (Fable) the session default caused built-in skills like /code-review and /simplify to inherit that model and exhaust the plan's usag…
Spain's data protection authority confirmed the country's first personal data breach caused by an autonomous AI agent, carried out by an individual using a large language model to attack an organization. The agency said …
A study titled "ImpossibleRubrics: Stress-Testing Generated Rubrics as Reward Signals" examines whether language model-generated rubrics used as reward signals for rubric-based reinforcement learning, LLM-as-a-judge eval…
A tutorial outlines a seven-stage learning roadmap for building voice agents, AI systems that communicate through spoken language via a three-stage pipeline of speech-to-text, language understanding and response, and tex…
A developer describes how repeatedly running AI code reviews on a small C# utility that deletes bin and obj folders caused the project to grow with each new model generation, as successive models found fresh critical iss…
Spain's data protection agency (AEPD) reported the country's first personal data breach caused by an autonomous AI agent, which used a known large language model to scan generic files, run vulnerability scans, and gain r…
Huawei supervisory board chairman Guo Ping told new employees at a September 10 internal forum that the company's goal in ICT and computing is "to be Nvidia," supporting any large model on its Ascend, Kunpeng, supernodes…
Google launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, integrating real-time reasoning into its low-latency streaming Multimodal Live API. The two production targets let conversational agents toggle deep-…
NVIDIA chief executive Jensen Huang rejected Anthropic co-founder Dario Amodei's call to slow frontier AI development during a joint appearance at Salesforce's Dreamforce conference in San Francisco on September 16, 2026…