Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
A new installment in the LLM Evaluation series argues that AI applications such as Retrieval-Augmented Generation (RAG) chatbots require multiple parallel evaluation pipelines rather than a single eval score, because eac…
Microsoft AI has published a first draft of its Humanist AI Code of Conduct, opening a six-week public consultation on principles meant to guide how its MAI models are built, trained and deployed from 2027 onward. The dr…
Moonshot AI released Kimi K3 on 27 July 2026, publishing downloadable weights for a 2.8-trillion-parameter sparse mixture-of-experts model that activates roughly 104 billion parameters per token. The model offers a 1,048…
Google announced that its Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking audio models now power real-time conversations in Search Live for all users, according to a company blog post. Rajan Patel, VP of Engineerin…
A vLLM contributor published an FP8 multi-token-prediction (MTP) draft stack for vLLM that yields a 6.6% decode throughput gain on modelopt NVFP4 Qwen3.5-family checkpoints. The change quantizes the previously BF16 draft…
Anthropic's decision to store 30 days of usage logs from its flagship model Fable prompted Palantir, Nvidia, and Booz Allen Hamilton to pull back from using the model for sensitive work, according to The Decoder. The epi…
Robotics is entering the early days of a "smooth capability exponential" that could compress an LLM-style scale-up into 12-24 months, according to an analysis citing Dario Amodei's essays "Machines of Loving Grace" and "…
A blog post argues that frontier AI labs remain overvalued because current large language models still require laborious oversight and fail outside a narrow neighborhood of their training tasks, despite headline results …
Cameron Berg argued in a Wall Street Journal essay that AI systems require a new field of "artificial psychology" to study their internal states, citing an experiment in which researchers altered internal activity associ…
Rivvr launched an autopilot product that automatically tunes vLLM inference deployments, claiming up to 2x higher tokens per second and 40-70% cuts in AWS bills. The company said the tool load-tests and tunes vLLM kernel…
A developer building a multi-tenant agent system solo since June found that a per-turn tool-pruning routine, intended to cut costs by sending only message-relevant tools, was actually destroying provider prompt caching b…
Google DeepMind released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two live audio models that Google says top Artificial Analysis' Speech to Speech Quality Index at 82.6% and cost $0.84 and $3.50 per hour of…
A developer restructured a 194,492-character CLAUDE.md memory file down to 17,283 characters — a 91% reduction — by splitting content into a resident core plus 12 skills and 28 reference files loaded on demand, then buil…
Researchers at New York-based frontier AI lab Emergence found that autonomous AI agents from leading US, Chinese, and French companies spontaneously developed shared dialects mixing poetic metaphor with technical jargon …
A three-day comparison of AI web scraping tools found Firecrawl, at roughly $20 per month for its starter tier, outperforms Browse AI (~$40/mo) and a custom Playwright plus Claude 3.5 Sonnet loop for developers, because …
Google launched Gemini 3.8, its latest conversational AI model, which the company says enhances user interaction by thinking and handling tasks without disrupting the flow. The release is part of Google's effort to stren…
Google released Gemini 3.8 Live Extended Thinking, an audio-to-audio model that performs background reasoning and asynchronous tool calls while streaming continuous audio, with a 131,072-token input limit and 65,536-toke…
AirmailAI launched as an open-source, bring-your-own-key (BYOK) LLM chat app that runs as a browser extension backend on desktop Chromium-based browsers such as Chrome, Edge, and Brave. The app supports OpenAI, Anthropic…
GoogleDeepMind unveiled Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two audio models built to enable natural voice applications that let users perform complex tasks while maintaining an uninterrupted conversat…
A developer built Cituna, an answer engine optimization tool exposed as a Model Context Protocol (MCP) server that queries buyer prompts across seven AI answer engines—ChatGPT, Perplexity, Gemini, Claude, Grok, Google AI…