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's API changelog lists 4 new response fields, all variants of input_tokens_details.cache_write_tokens_details across ResponseUsage, BetaResponseUsage, BetaResponse.usage and Response.usage, alongside 2 breaking cha…
Anthropic released FlashProve, an Apache 2.0 zkSNARK prototype that cryptographically proves a reply came from a committed model without revealing its weights. Anthropic reports FlashProve proves Llama 4 Maverick at 110 …
A developer-built comparison of three LangGraph research workflows found that splitting research, writing, and critique into fixed roles raised completion from 8 of 15 runs for a single ReAct-style agent to 13 of 15, wit…
Frontier AI models can now generate CAD, design custom PCBs, source commercial-off-the-shelf parts and describe assembly, but the lossy interface between digital designs and physical atoms means superhuman hardware engin…
A developer released Claude Code Local, an MIT-licensed Python server that lets Anthropic's Claude Code run against open models on Apple Silicon Macs with no API key, cloud dependency or usage limit. The server listens o…
Simon Willison released llm 0.37.1 on 3rd November 2026, a bugfix release of his command-line tool and Python library for running large language models, with contributions from Steven Huang, Tao An, Arjan Mossel and XQ. …
MoEShift, an alpha-release local inference runtime from developer Davide Troiani, runs a 125-billion-parameter Qwen 3.8 Flash Next model on Apple Silicon at up to 11 tokens per second with under 10 GB of peak process mem…
A task-by-task comparison finds Anthropic's Claude Opus 5.5 is the better default for most day-to-day coding in late 2026, priced roughly 2.5x lower per token than OpenAI's GPT-6 Astra, while Astra at its high or max eff…
ChatGPT handles roughly one-fifth as many search-like queries as Google, according to ChatGPT consumer product lead Adam Fry, who said OpenAI counts a search each time a question sends ChatGPT to the web. Fry gave the co…
A developer released Grai, a single-file offline HTML writing tool (grai.html) that extracts a writer's "Voice DNA" from 5 to 20 past posts and audits drafts against 6 structural rules and a CV ≥ 0.32 rhythm floor, runni…
AmpUp launched Enterprise Brain, a private-preview context layer for enterprise agents that stores company facts in a vertical-specific schema and rewrites task notes after every graded run, claiming 35.2% root-cause acc…
NVIDIA demonstrated three stacked optimizations that raised GSM8K decode throughput 42.5% on Ant Group inclusionAI's 124B Ling-3.0-flash model running MoE inference on a single DGX Spark, with no detectable accuracy chan…
Anthropic's Project Glasswing, launched April 7th, 2026 with access to the unreleased frontier model Claude Mythos Preview, had by May 22nd produced more than 10,000 high- or critical-severity software flaws across about…
Developer lucifix published NewsNest, a self-hosted RSS and Atom news reader that uses Anthropic's Claude API to score each story on topic, importance and personal fit against a user-written interests file, with a defaul…
Anthropic released Message Threads in public beta on its Messages API as of November 2, 2026, letting Claude agents send only the new message per turn while the API caches the prior conversation server-side, which Anthro…
Revolut cut its AI coding inference costs by roughly 90 percent in under a year by making self-hosted GLM models its default instead of commercial services like Anthropic's Claude, David Abitbol, Revolut's ops and infras…
MIT CSAIL researchers reported in the paper "Sparks of Weight-Space Programming" that Astra, a weight analyst built on OpenAI's GPT-6, raised the MATH-500 score of the 800-million-parameter Qwen3.5-0.8B from 7.9% to 48.1…
A research team has built a system that formalizes Nvidia's PTX semantics in Lean and uses an LLM to write Lean proofs that GPU kernels compute the correct result, with Lean's small, well-audited proof checker verifying …
Elon Musk said the cost of running Grok Bot, the personal AI assistant SpaceX introduced in August, is "unsustainably high relative to revenue," and OpenAI consumer head Adam Fry said OpenAI's Dots assistants are "on the…
SemiAnalysis's GDPx benchmark, released November 2 and built with 14 specialist partners across 206 long tasks in 11 high-value domains, found the best model, Claude Opus 5.5, scores 12.0%, with Gemini 4 Carbon at 10.9% …