Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Google Threat Intelligence Group (GTIG) warned that financially motivated "LLMjacking" campaigns are hijacking enterprise cloud and LLM credentials and reselling model access through underground subscription services, ge…
Researchers from South Korea and the US found that changing the activation of a single neuron in each of five open-weight models — Qwen3, Llama 4, DeepSeek R1, Gemma 4 and Mistral Small — could flip investment recommenda…
An ongoing open-weight LLM benchmark project, BudHist, reports that qwen3.6:35b leads tested local open-weight models on Buddhist history multiple-choice questions with 210/250 (84%), while cloud-based open-weight models…
Vishesh Tripathi posted an arXiv paper on 18 June 2026 (revised 23 June 2026) proposing Grouped Query Experts (GQE), a mixture-of-experts layer built on grouped-query attention (GQA) in which a router selects k query-hea…
Five AI models — ChatGPT, DeepSeek, Grok, Gemini and Nemotron — all chose the same option when asked which of three retirement packages a Chinese servicemember with 18 years of service should take: 10,000 RMB per month u…
Open-source desktop app Strata lets users run AI models locally for chat, coding and image work without an account or an expensive GPU, serving models through local OpenAI- and Anthropic-compatible APIs that support Clau…
ML Mastery published a tutorial on selecting an LLM inference optimization technique by working through a decision tree that matches the method to the actual bottleneck in a workload. The article's available text does no…
Google's Mandiant business has identified a threat it calls LLMjacking, in which attackers hijack foundation models to run up fraudulent charges exceeding $100,000 in a single day, with some organizations losing as much …
OpenAI quadrupled bug bounty payouts, raising rewards for critical vulnerabilities in products such as ChatGPT and Codex to up to $30,000, with exceptional findings reaching $100,000 and roughly $172,000 including qualit…
Trim launched as an OpenAI-compatible local proxy at http://127.0.0.1:8888/v1 that compresses coding prompts before they reach upstream LLM providers, aimed at clients including Cursor, Continue, VS Code Chat and Claude …
A developer has announced Dagento, a self-built local AI agent framework written in PHP for Linux, intended to address setup difficulty and reliability problems the developer encountered in Hermes Agent, OpenClaw and Age…
A tutorial published by Shattered outlines a 12-step, 75-minute approach to building a layered guardrail gateway in standard-library Python that sits in front of any chat model, normalizing input, screening for injection…
A free, open-source course called AI Red Team Academy has been released, offering 60–80 hours of offensive security training across 8 modules on attacking LLMs, RAG pipelines, multi-agent systems, and ML supply chains. E…
Artificial Analysis's overhauled Intelligence Index v5 widened the gap between the best open-weight AI models and the overall leader from 3 points to 12 points, with open-weight leaders scoring 32 against 58 for Anthropi…
Revolut moved roughly 2,000 engineers onto self-hosted GLM for coding and placed Claude Code and Codex on an exception list, according to David Abitbol, who leads AI operations at the company. Abitbol said the setup runs…
An OpenAI model, Astra, running through OpenAI's Codex environment, scored 49 out of 50 (98%) on the 50-problem TheoPhysEval theoretical physics benchmark after researchers from Harvard, the University of Texas and Googl…
Anthropic released FlashProve on November 2, 2026, an open-source research prototype built on a hash-based zkSNARK that verifies which model produced an output without revealing its weights, activations or routing decisi…
Tokely mapped the 21 evaluations on Artificial Analysis's Intelligence Benchmarking methodology page (as of November 3, 2026) to the 41 Generalized Work Activities in O*NET 31.0 and found that the composite Intelligence …
MIT CSAIL researchers Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab showed that OpenAI's GPT-6 Astra can directly edit another model's weights, changing only 0.016 percent of Qwen3.5-0.8B's weights to improve i…
Andrej Karpathy built HackerTube, a YouTube-style site that surfaces only the 6,454 videos Hacker News commenters recommended between 2007 and 2026, with Claude Opus 5.5 doing nearly all the work in about two hours. Karp…