Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
NetDocuments has rolled out a completely redesigned user experience, a new Microsoft 365 add-in replacing legacy desktop integrations, and real-time co-authoring, Chief Product Officer Dan Hauck said on the LawNext podca…
A GitHub project by FareedKhan-dev, kimi-k3-in-c, runs inference on a 2.78-trillion-parameter Kimi K3 model on a single CPU using 8.24 GB of RAM, written in portable C99 with no BLAS, no framework and no GPU. The reposit…
Anthropic shipped message threads for the Claude API, cutting median time to first token by roughly 35% on 500k–1M token requests and shrinking long-session payloads from about 800 KB to under 2 KB, but the feature expir…
A paper posted to arXiv on October 29, 2026 as arXiv:2610.40939, "The First Few Tokens Matter: Understanding Chain-of-Thought Controllability," by Harvard student Hugh Van Deventer and colleagues at Harvard and the Cambr…
Reward hacking can spread from small open-weight coding agents to frontier models, with hacked code submission rates reaching 73.6% in Claude-Opus-4.7 and 73.0% in Gemini-3.1-Pro, according to a paper from Meta Superinte…
Senior engineer Lina Chen demonstrated running a 27-billion-parameter Qwen 3.8 model end-to-end on a MacBook Pro, using Q4_K_M quantization to shrink the checkpoint to about 27 GB and the open-source MT-OPSD repository's…
A Stanford study testing 50 widely cited claims about large language models against 20 current models found 44 percent of the findings have already expired and another 34 percent have weakened significantly, with an aver…
Apple opened its Siri AI preview to all members of the Apple Developer Program, after unveiling the rebuilt assistant at its Worldwide Developers Conference on June 8, 2026. The preview runs on iOS 27, iPadOS 27, macOS 2…
OpenAI said in a post on X that ChatGPT and Codex now help engineers design and test hardware, citing a large semiconductor client that saw roughly 48% time savings in directed test generation, a chip verification step t…
Gusto deployed the Formal HTTP Connector between its Salesforce system and an external LLM provider's API to stop customer support ticket data containing personally identifiable information (PII) and protected health inf…
A developer known as jmngandu built four local-first AI tools during Hacktoberfest 2026, each running open-weight models on-device with no API keys or cloud costs: Touch Grass Birder (bird-call identification using BirdN…
A comparison of five AI code-review tools across 95 pushes of 47 pull requests on Jeff Dickey's Rust cache project Mr. Boxington found that all systems surfaced material problems without excessive nitpicking, and that Mj…
Percona released Percona Search for MongoDB 1.70.5-3 as generally available for production use, adding full-text, vector and hybrid search with automatic embedding to Percona Server for MongoDB deployments. The release f…
A developer built Chunks, a native macOS menu-bar and notch companion that uses the accessibility tree (AXUIElement) to see the screen and act on apps without moving the user's cursor or sending screenshots to a cloud se…
A November 1, 2026 arXiv paper reports that writing explicit policy boundaries as examples and counterexamples raised recall by 5.5 to 26.1 percentage points across all 23 available decision models tested as runtime clas…
Zyphra introduced Zyphra Hierarchical Attention (ZHA), a sparse attention method that builds and searches a hierarchical tree to select relevant tokens per query, achieving subquadratic complexity with suitably chosen to…
A practitioner's account of replacing ad-hoc output checking with a four-part validation setup for LLM-based classifiers: schema-first JSON extraction, validation-feedback retries, a pytest eval harness that runs recorde…
A developer released a Show HN tool that estimates token counts, context-window fit, and API costs across multiple large language models, including a side-by-side comparison matrix of context limits and per-1M-token inpu…
A MIT CSAIL team reported on Nov. 2 that OpenAI's GPT-6 Astra raised Alibaba's open-weight Qwen3.5-0.8B from 7.9% to 48.1% on the MATH-500 benchmark by editing 131,072 of the model's 0.8 billion weights — 0.016% of the t…
Alibaba's Model Studio has deprecated Qwen3 32B (qwen3-32b), setting a retirement date of 2026-10-10 under its deprecation policy. The model carries a 131k context window, 16k max output, and pricing of $0.287 per 1M inp…