Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Nvidia is preparing to release a free AI model built by Poolside, the U.S. coding startup in which Nvidia spent $7 billion last month to buy a stake and license its model-building tools, with the model due by the end of …
A Google-led research paper titled "Procedural Graphs: Self-Evolving Execution Structures for LLM Agents," submitted to arXiv on September 8, reports that a framework organizing procedural knowledge into editable graph s…
A study titled 'The pain axis: LLMs represent self-directed harm and act to relieve it' found that all 25 open-weight AI models tested responded to an activated "pain vector," pressing a relief button in 25 to 71 per cen…
Chinese AI lab StepFun released Step 5 Preview, a flagship agentic model that scored 44 on the Artificial Analysis Intelligence Index, beating Google's Gemini 3.8 Flash (high) at 41 while costing $0.72 per Intelligence I…
On Terminal-Bench 4.0, the benchmark co-hosted by Stanford, Harbor, and the Laude Institute for multi-step software tasks, the best Chinese entry is Zhipu's GLM-5.3 at 41.8% as of September 19, while OpenAI's GPT-6 Astra…
LanguageOps has built automated checks into its translation editor to catch mechanical formatting errors such as misformatted placeholders like `$0` and `{name}` before reviewers see them, according to the company's guid…
InfoWorld published a list of 20 prompt styles and structures for getting better results from large language models, including step-by-step instructions, few-shot examples, Q&A-style questioning, and prompts that ask the…
Thomson Reuters released Thomson-1.0-Small, a 35.1 billion parameter Mixture-of-Experts causal language model built on the Qwen3.6-35B-A3B architecture with a 262,144-token native context length, developed in partnership…
A developer built a Data Analyst Agent using Google's open-source Agent Development Kit (ADK) and the Gemini 2.5 Flash model, demonstrating agentic workflows over single-prompt chat completion. The build hit two configur…
Yandex released AliceAI-Foundation-80B-A3B-Base, an 80B-parameter mixture-of-experts language model trained entirely from scratch on roughly 18T tokens and published on Hugging Face under Apache 2.0. The base model activ…
A developer published jev-leftpad, an npm package that left-pads a string by calling the jev-latest model through TypeSafe's @typesafe-ai/sdk instead of JavaScript's built-in padStart(). The package requires Node.js 20 o…
A developer benchmarked whether large language models can judge UI style-rule compliance from screenshots by comparing GPT and Gemini verdicts against DOM-derived gold labels across 60 contrastive rule pairs. GPT scored …
Developer Manuel Cecchetto released codex-context-gc, an unofficial tool that exposes a compact_context function letting OpenAI's Codex desktop app compact its own context after a verified phase and continue work automat…
Jared Palmer, vice president of engineering at Cognition, released a Qwen3.5-based generation of the open-source Kev decision-model project on September 20th, comprising 0.8B, 4B and 9B parameter variants that return typ…
A new non-profit site, pressaudit.org, is using Alibaba's Qwen 3-32b large language model to process thousands of news articles and analyze them by region, public interest, and partisan lean, according to the project's l…
InclusionAI released Ling-3.0-Flash-Sante, a Mixture-of-Experts model built on Ling-3.0-Flash with 124 billion total parameters that activates approximately 5.1 billion parameters per token. The model targets health and …
Jared Palmer released Kev, a family of small decision models built on Qwen3.5 in 0.8B, 4B, and 9B sizes, with pretrained weights, training code, and evaluation data available on GitHub. Kev's API matches TypeSafe's Syste…
A developer built a ~10.6 million parameter decoder-only Transformer from scratch in PyTorch and trained it end-to-end on the Tiny Shakespeare dataset using only the free tier of Google Colab's T4 GPU. The 6-layer, 6-hea…
A developer reported that a local model, Ternary Bonsai 2, generated the parse_prompt function returning ParsedPromptInfo from a 180-line Markdown specification in about 30 minutes, producing 300-400 lines of Rust with 3…
A benchmark of 50 product-classification tasks found that replacing an agentic LLM pipeline with TypeSafe's System One model, Jev, cut mean latency 86% from 9.62 seconds to 1.38 seconds and reduced model calls 56% from 7…