Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Apple opened Siri AI to all third-party developers on November 3 through the iOS 27.3 beta, letting any app expose actions to Siri by adding the single Swift macro @AppIntent(.tool) to an existing AppIntent struct, which…
The EU AI Act (Regulation (EU) 2024/1689), which entered into force on 1 August 2024, is in the third of five implementation waves as of 4 November 2026, with prohibitions and AI-literacy provisions applying since 2 Febr…
ML Mastery published a tutorial on turning an open-weight language model into a decision model that classifies user intent by fine-tuning Qwen3.5-4B with Unsloth. The tutorial is aimed at developers who want to build the…
A developer built GhostShell, a roughly 30-line Python terminal wrapper that runs each shell command via subprocess and then sends the command and its output to a local Ollama server running llama3.2:1b with a ghost pers…
A Tsinghua University preprint released around October 14, 2026 found that no tested AI agent memory system successfully handled more than 1.5% of user forget requests end to end, with most scoring 0%, using a new benchm…
A team's automated code-review system, built around a self-hosted 27B model with transparent fallback to hosted providers such as Anthropic, delivered feedback on 75% of pull requests within 8 minutes during its first mo…
An independent replication by blogger flecart confirmed that the MIT lab's weight programming results hold: the Astra model achieves a high score on Breakout by compiling a roughly 10-line program directly into network w…
Mistral AI released a demonstration on November 4, 2026 in which its open-weight Mistral Large 4 model used 133 specialized agents to triage, reproduce and patch a recently disclosed Django vulnerability, one day after D…
A developer's open-source harness, KYNEX, scored 92.97 F1 on LlamaIndex's ExtractBench (370 documents, 4,869 pages), ranking 12th of 53 systems and beating GPT-6 Astra (91.9) and Claude Code with Opus (87.1) while runnin…
OpenRouter listed a free model called Crouton Alpha under its anonymous "Stealth" provider tag at 01:38 UTC on November 3, offering a 1,048,576-token context window, up to 944,000 tokens of output, and text, image, audio…
A developer published a step-by-step guide showing how to stream Claude tokens to the browser using a minimal Fastify service on AWS Lambda, calling Bedrock through Node.js 22's native fetch and forwarding tokens as Serv…
Prompts.chat has turned natural-language prompts into a version-controlled, open-source corpus, converting static prompt engineering into an MDX-driven metaprogramming pipeline and agent graph, according to the project's…
Anthropic announced on October 30, 2026 that fast mode on Claude Opus 4.8 is deprecated, and from November 30, 2026 any request to `claude-opus-4-8` sending `speed: "fast"` will return an error rather than silently downg…
A workshop at the AI Engineer Code Summit in San Francisco on 10 November 2026 will teach attendees to build a coding agent harness in TypeScript on Claude Haiku 5.5, starting from a bare `while (true)` loop around a sin…
A new blog post outlines testing strategies for randomized algorithms, using the AWRS sampling algorithm from Lipkin et al. (2025, arXiv:2504.05410) as a worked example for constrained generation in sequential Monte Carl…
Tencent released Hy4, a 770-billion-parameter mixture-of-experts flagship in its Hunyuan line, marking the first model in the family to combine fine-grained image understanding with computer use through graphical interfa…
An analysis of benchmark evidence finds that open-weight models such as Qwen3-Coder and DeepSeek now perform within roughly 2.2% to 6.4% of proprietary models on SWE-bench Verified when run inside the OpenHands coding ag…
Linux on Nvidia RTX Spark laptops such as the MSI Prestige N16 Flip AI Plus locks the GPU to roughly 780 MHz when plugged in, versus about 2.2 GHz on Windows, because the Nvidia Platform Controllers and Framework power-m…
An audit of public LLM evaluations found that replacing prose request descriptions with JSON Schemas and loosening unstated grader requirements sharply changed scores on Zapier's AutomationBench, with Claude Opus 5.5 @ m…
Placing a heart emoji next to a coding problem caused Claude and GPT to write recursive code 100% of the time versus a 0% baseline, according to research by Alexa R. Tartaglini conducted during a summer at Google DeepMin…