Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
DeepSeek added a Creator Mode to its open-source DeepSeek Harness that generates plugins directly from text prompts, the company announced. DeepSeek demonstrated AI-generated plugins including a digital sticky note wall,…
Zyphra AI introduced Zyphra Hierarchical Attention (ZHA), a tree-indexed sparse attention method that retrofits frozen LLMs using 4B tokens of distillation. Zyphra reports ZHA matches dense attention on reasoning and 256…
Researchers at MIT CSAIL — Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab — published "Sparks of Weight-Space Programming," showing that GPT-6 Astra can read and edit raw neural network weights without gradient…
An independent developer released Wisteria, a BYOK desktop chatbot that stores conversation logs, persona memories and compressed long-term memory as text files in a local directory on the user's device. Wisteria uses a …
Anthropic has tightened its usage rules for Claude to prohibit using the model to build copycat games or reverse-engineer code, Boing Boing reported on November 2, 2026. The change targets a wave of AI-generated game clo…
Hippocratic AI co-founder and CEO Munjal Shah announced Polaris 6 Ultra on November 2nd, a healthcare language model the company says reads a patient's emotional state and plans its next sentence mid-speech, a capability…
AIClaw 3.7.32, released on GitHub, adds service-aware routing that resolves a model's connection contract from the configured provider type before choosing a protocol client, rather than inferring the interface from a mo…
A paper titled Latent Sparse Attention (arXiv 2610.42147), submitted October 30, 2026 by Ales Kubicek, Torsten Hoefler and eight co-authors, reports that its D-LSA variant speeds up a single attention layer by 22.7x in p…
Simon Willison released LLM 0.37 on November 2, 2026, adding support for OpenAI's GPT-6.1 Sol and GPT-6.1 Terra models plus a `chat-latest` alias, and switching the tool's default model to GPT-6 Luna, which Willison says…
Anthropic released FlashProve, an open-source Apache 2.0 system that proves an AI response was produced by one committed set of model weights without revealing them, generating proofs on Llama 4 Maverick (a 400B-paramete…
Hugging Face proposed Model Integration Contracts (MICs), a standard for packaging small "fixture" models with runnable usage paths and human-reviewed baselines, published in the huggingface-mic repository, after a July …
A study by Owen Queen and James Zou of the Zou group retested 50 widely cited LLM findings across 20 models released from July 2024 to September 2026 and found that 44% (22 claims) had expired by September 30, 2026, 34% …
OpenRUA released uni-robot-agent, an MIT-licensed open harness that lets language models operate robots through tool calls, and its open-weight Qwen3.8-Flash-Next model scored 31.39 on the RoboDojo benchmark, above the p…
In tests on 21 live product pages from Amazon, Walmart, Best Buy, Target and four Shopify stores, GPT-6 Luna and Claude Opus 5.5 each returned the exact live price for only 2 of 15 products when answering from memory, wi…
Gartner senior director Mark Petty and group vice president Raymond Paquet argue in a paper, "Open-Weight Models Will Run 90% of Enterprise Agentic Workflow Steps," that open-weight models cost a quarter as much as the c…
Google's Firebase iOS SDK 13.0.0, detailed in the Firebase Apple SDK release notes for Oct. 7, 2026, requires Xcode 26.2 or later, Swift tools version 6.2.1 and the Swift 6.2.3 compiler, and drops CocoaPods in favor of S…
An AI-assisted review by Epoch AI, the group behind the Frontier Math benchmark, flagged fatal errors in roughly a third of the benchmark's problems, with the team judging most of those flags valid. Separately, a team in…
A frontier AI model scored 49 out of 50 on TheoPhysEval, a 50-problem unpublished physics benchmark built by researchers from Harvard, UT, Google DeepMind and other institutes, after physicists rechecked their own answer…
ISTA DASH Lab released DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUF, a compressed build of DeepSeek V4.1 Flash that prunes roughly half of the base mixture-of-experts layers and quantizes the survivors to an effective 1 b…
NVIDIA's Nemotron Labs released Dory, a 3.97 billion parameter looped transformer that reuses a 40-layer recurrent block up to eight times per token and beats Gemma 3 27B on a composite benchmark score despite far fewer …