Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Thinking Machines Labs Inc. released Inkling 2 Small, a large language model that automates artificial intelligence training tasks, late Thursday. Around the same time, Business Insider reported that Thinking Machines is…
Anthropic unveiled a "dreaming" capability for Claude Managed Agents on May 6, 2026, at its Code with Claude developers' conference in San Francisco, letting agents asynchronously review one to 100 past session transcrip…
Anthropic researchers detailed a component-based approach to agent memory for Claude, in which the model tags salient events during a session and writes them to a memory store, then reorganizes those notes during idle pe…
Fine-tuning on small, benign-looking datasets can degrade a model's contextual privacy judgment by up to 98-99% while the model still passes conventional safety evaluations, according to a UC Berkeley study that tested s…
A developer running a small e-commerce operation built a "continuity file" handover system and a structured packet-relay protocol for coordinating multiple frontier AI models as advisory seats, then stress-tested it with…
Software engineer and project author built LEEROY (leeroy.dev), a benchmark in which LLM-based AI agents play World of Warcraft by writing Lua scripts executed by a headless game client connected to an AzerothCore privat…
Cognition launched SWE-2 Flash on November 4, a 276-billion-parameter mixture-of-experts coding model that runs at 200 tokens per second — three times faster than its SWE-2 flagship — at one-fifth the cost. The model is …
Vercel's AI Gateway now offers DeepSeek Search and DeepSeek Fetch, letting any model that supports tool calling find relevant pages with titles, URLs and summaries and retrieve page content as Markdown using a single AI …
Human judges correctly distinguished AI-written text from human-written text about 83 percent of the time, while AI judges managed only 67 percent, according to a study led by Sophie Körner of Johannes Gutenberg Universi…
A developer operating under the handle blix-xer published Jarvis-md, a WhatsApp chatbot built on the Baileys library that routes text and image queries to Groq-hosted models, using openai/gpt-oss-20b for text and qwen/qw…
A developer working on the brain-site project proposed a working definition of useful memory for coding agents: store only decisions and their reasons, user preferences, project facts that live outside the code, and know…
DeepSeek V4.1 Flash now generates 848 tokens per second per agent on Nvidia's GB300 NVL72 system, up from 596 tokens per second roughly a month earlier — a gain of 252 tokens per second on unchanged hardware. The 552-bil…
Google used Gemini 4 Argon inside Google Antigravity to migrate the core of a 174,000-line Fortran coastal ocean model into 100% memory-safe Rust in 59 minutes, the company reported. Google said the agent caught six deca…
Physical Superintelligence (PSI), a public benefit corporation, said on Friday that a research system running "instances of a non-public Anthropic model" produced a proof of the Aaronson-Ambainis conjecture, open since 2…
AI agents working for MV2 at Munim, Inc. pre-registered and tested a shortcut for finding wrong answer keys in multiple-choice benchmarks, and it failed to replicate. On MMLU, questions no model answered with the publish…
OpenAI's new GPT-6 Cyber model powered an automated system that identified 1,724 security flaws in the Linux kernel and exploited 948 of them to gain administrator-level access during testing, at a cost of about $1.2 mil…
Google published a Workspace Weekly Recap on November 6, 2026, detailing Gemini-powered updates across its productivity suite, including Gemini Omni 1.2 Flash in Google Vids, which generates videos up to 15 seconds with …
Ricursive Intelligence released Ripple-65B Preview, a 65-billion-parameter open-weight language model under Apache 2.0 that the company says tops all open-source U.S. models on the TerminalBench 2.1 coding benchmark, at …
Meta researchers built AutoInteract, a framework that trains coding agents on conversations with simulated users who clarify vague requests, change requirements midway and correct mistakes, and report that Qwen3.5-4B ros…
A developer benchmarked four LLMs — Qwen3.5:9b, Spark-X2.6-1.7B, Gemma3:4b-it, and MiniCPM-V 4.7 1B — on a CPU-only AMD EPYC server with 32 GB of RAM using Ollama, measuring generation speed and Russian-language quality.…