Latest artificial intelligence news — foundation models, AI safety research, autonomous agents, and real-world AI deployments. Updated every 30 minutes from top AI publications.
Andrej Karpathy proposed an approach to knowledge management called LLM Wiki, in which an LLM maintains a persistent, interconnected base of Markdown pages instead of re-retrieving raw documents for every query. The arch…
A developer built Haunted Terminal, a zero-dependency Python tool that reads shell history across PowerShell, Bash, Zsh and Fish and pipes recent commands into an open-weight model (llama3.2:3b) running locally via Ollam…
NVIDIA's Nemotron 3.5 Super VL, a 120B-parameter mixture-of-experts model that activates 12B parameters per token, is now available on AWS alongside the existing Nemotron 3 Nano, Nano Omni, Super, Ultra and Nemotron 3.5 …
Developer Yankel Perelman released Anna, a developer-edition voice assistant that runs on a user's own machine or server with their own Gemini API key and reports the host's CPU load, RAM, disk and uptime by voice throug…
Google's AI Mode search tool returned no answer for much of Wednesday when asked who won the 2020 presidential election, while it correctly reported the 2024 result, and Google attributed the non-answers to "a technical …
Google announced general availability of expanded semantic model support in Knowledge Catalog, which now natively ingests metric definitions from Looker LookML, BigQuery Graph (including BigQuery measures), dbt MetricFlo…
OpenAI has expanded its misalignment monitoring system to cover frontier models across training, evaluation, internal use and external deployment, after the Hugging Face incident began during internal cybersecurity evalu…
Cisco Talos detected malware that uses stolen OpenAI API keys to run AI queries through compromised machines, with one variant capable of executing over 52 commands, leaving victims to foot the bill. The Go-based implant…
Mistral released the open weights of Mistral Large 4, codenamed Le Chonk, a 1-trillion-parameter natively multimodal mixture-of-experts model aimed at coding, cybersecurity and enterprise agents. Mistral reported the mod…
Meta Platforms Inc. pledged on Wednesday to consult news industry groups before finalizing new terms of service after publishers objected that the terms could let Meta use their content to train its artificial intelligen…
A developer describes building multi-agent systems that repeatedly failed the same way: agents communicated their decisions and reasoning but not their flexibility, so individually correct agents converged on collectivel…
Only two of 39 large language models tested on RubiksBench could reliably solve a Rubik's cube without tools, with most models unable to solve a cube just six moves from completion, according to the benchmark's single-re…
The team behind Plant Finder rebuilt its plant-identification app around diagnosis rather than naming, after testing with thousands of gardeners showed users mainly want to know why leaves are yellowing or spotting. The …
A developer built GhostMemo, a browser-based memory journal that stores entries in local storage and uses a locally running Ollama instance with the Gemma 3 1B open-weight model to turn saved memories into fictional ghos…
A GitHub project by developer yeet-masta, Nutty-Fluffies-Rollercoaster-Decomp, uses ILSpy plus Claude Opus 5.5 to clean up decompiled code from the game Nutty Fluffies Rollercoaster, renaming roughly 600 local variables,…
Google unveiled Gemini 4 Carbon on Wednesday, a frontier model that Koray Kavukcuoglu, senior vice president of Google DeepMind and chief AI architect at Google, said leads SEC-bench Pro at 88.2% for finding vulnerabilit…
Paris-based Mistral released the weights for Large 4, nicknamed "Le Chonk," a trillion-parameter model trained entirely in Europe. Mistral says Large 4 beats its open rivals at coding, and independent testing shows it is…
Mistral AI SAS released open weights for Mistral Large 4, nicknamed "Le Chonk," a 1-trillion-parameter natively multimodal model with 52 billion active parameters that was first shown in preview three weeks earlier. Mist…
Linearbytes released Bug0 QA Agent, a source-available Playwright library for agentic browser regression testing that caches successful single-action steps in an in-memory Redis store and replays them directly through Pl…
A developer built PHANTOM, a local-first AI security investigator that pairs deterministic code scanning with locally running language models via Ollama to explain suspicious findings and suggest what to investigate next…