Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
OpenAI released GPT-6 Cyber on November 5, 2026, a reasoning model built for software vulnerability detection and security auditing, available initially to select Daybreak Red customers with access expanding over time. O…
A developer's workflow splits coding between a frontier model acting as an "architect" that designs and reviews but never writes code, and a smaller, cheaper model acting as the "engineer" that does the implementation, c…
Anthropic researchers analyzed 207,000 Claude and Claude Code research sessions using privacy-preserving methods and found that data analysis accounts for about one-third of sessions on Claude and nearly half on Claude C…
The Draw Things team released Local Code, a free, open-source AI agent for Mac, on October 28, 2026, following a public beta that began September 12. The desktop app runs Qwen or DeepSeek models entirely on-device, bundl…
A developer built ctx 1.3.0, a TypeScript command-line tool published on npm as yodaman-ctx, that indexes local codebases into SQLite and retrieves code using BM25 and dense vectors fused with reciprocal rank fusion, ans…
NVIDIA Research and Inferact's REINFORCE Zero trains stably for 20,000 updates at policy lag 64 and reaches 2.40× the training-step throughput of synchronous reinforcement learning once rollout and training kernels are a…
A Hugging Face repository named TheSanchezian/Nemotron-3.5-Lightning-30B-A3B-Ternary-OpenVINO-A770 has been listed in the platform's model catalog, indicating a ternary-quantized Nemotron 3.5 Lightning 30B A3B variant pa…
A security lecturer backdoored an open-weight code model with roughly ten poisoned examples for under £75 in under an hour, while Anthropic research showed the number of poisoned documents needed to backdoor a model stay…
GateBolt, a two-layer product for securing AI agents, cut prompt-injection password leaks to near zero at its most cautious setting in a test on a door-keeper chatbot called Hodor, but still let AI attackers win 16.6% of…
A developer recompiled David Horner's 1997 Windows 3.11 game Drain Storm for modern systems using multiple GPT models in a "vibeslopping" decompilation, with the first pass producing a completely non-functional build and…
A production agent study found that delivering table rows over a typed side channel instead of the model's context raised fully correct turns from 45.7% to 92.3%, one of seven arXiv papers on agent failures reviewed in n…
Researchers from Anthropic's Fellows Program published a multi-agent welfare study on the Alignment Science Blog on November 5, 2026, testing five models — Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro, Grok 4.3, and Kimi K3 …
A developer benchmarked gpt-image-2.6's native output_format: "svg" support while building GPTSVG Studio, finding that native LLM-authored SVG geometry produces compact files (0.9 KB to 9.6 KB, 5 to 28 paths) with groupe…
A developer who built a deep research agent that searches the web, browses sites, gathers evidence from dozens of sources and produces cited long-form reports reports that most engineering difficulty came from context ma…
Thinking Machines Lab released Inkling 2 Small, a 594-billion-parameter model with 17 billion active parameters, which scored 31 on the Artificial Analysis Intelligence Index v5.0.1 — the top score among models built out…
Zml/llmd released version 20261106.0, adding support for Reflection's Beam 1.0 (501B-A23B) and Day 1 support for Ricursive's Ripple (65B-Preview), plus beta asynchronous scheduling for Llama and Gemma 4. The asynchronous…
Google released a new `.bigquery` accessor in its BigQuery DataFrames (BigFrames) library that lets pandas users call BigQuery AI functions such as `ai.predict()`, `ai.embed()` and `ai.similarity()` directly on DataFrame…
Tencent said on November 5 that it expanded inference capacity for its Hy4 model tenfold after heavy first-day demand on WorkBuddy pushed many users to their usage limits, and the company raised free usage allowances. Hy…
Netflix staff research scientist Cameron Wolfe presented a framework for evaluating AI on tasks without a single correct answer, arguing that LLM judges are random and that resampling verdicts shrinks the error bar. Appl…
A new Behavior-Conditioned Symbolic Synthesis (BCSS) framework, detailed in an arXiv paper (2611.07088), automatically extracts deterministic, interpretable symbolic programs that verifiably reproduce the reasoning pathw…