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, a former OpenAI founding team member and Tesla Autopilot AI lead who joined Anthropic earlier this year, built HackerTube, a video site drawing on 6,454 YouTube videos recommended across roughly 43 milli…
Roe AI released nonce, an open-source tool on GitHub that brokers least-privilege Gmail, Calendar and Drive access for personal AI agents including Muse, Grok, ChatGPT and Claude, so agents no longer hold the user's Goog…
Developer Rob Cox has published a free series of 18 from-scratch Python projects, aimed at curious people aged 12 and up, that let readers build working versions of an operating system (PyOS), a programming language (Peb…
A reverse-engineering project published an open replacement firmware for the Fosi Audio MD3 USB-C DAC and headphone amplifier, which uses an ESP32-S3R8 microcontroller and a Savitech SA9312L audio bridge. The replacement…
Developer Nick Bild built a generative image system that runs on a Commodore 64, adapting a probabilistic PCA algorithm to produce 8x8 retro game sprites from roughly 100 hand-made training sprites. Bild trained the mode…
DeepSeek released dsh-superagent on October 31, an open-source, MIT-licensed coding agent that monitors GitHub issues and Feishu conversations and autonomously submits pull requests without being prompted. The agent uses…
Anthropic added two sentences to Claude Code's prompt on Opus 5.5 to steer the coding agent away from recreating existing games, Anthropic member of technical staff Addy Osmani said on X, stating that "using Claude to in…
A developer released Purser, an MIT-licensed ERC-4337 smart account that lets AI agents hold and spend funds through session keys governed by on-chain policies rather than raw private keys. Each agent key carries per-pur…
An intern at RakFort fact-checked every finding produced by Secfoo, an open-source tool that turns AI coding agents like Claude Code and Codex into security reviewers, against the tool's own codebase. Six of seven findin…
A developer writing on Chain of Thought argues that applications should let language models abstain rather than guess, citing Scaled Cognition CTO Dan Klein's claim that enterprise audits reveal real hallucination rates …
Anthropic has obtained an O(n^1.9992) algorithm for 3SUM, breaking the long-held conjecture that the problem has no subquadratic algorithm, according to a blog post that assumes recent results by OpenAI and Anthropic con…
Instinct, the personal AI assistant built by San Francisco startup Spear Street Technology, stopped sending WhatsApp messages for multiple users, who reported the outage on X and said the agent attributed it to a billing…
The Linux 7.4 merge window drew 19,671 Git commits since Linux 7.3, the busiest merge window in kernel history, according to Phoronix. The kernel source tree reached 41,605,120 lines ahead of v7.4-rc1, including 31,417,9…
Judge Luiz Carlos de Araujo Santos Junior fined two Brazilian lawyers R$84,000 (about $16,000) and referred them to the state bar after they hid white-on-white instructions in a May court filing telling the judicial AI s…
An open-source project called AiSOC (6,106 GitHub stars, MIT licensed) provides a self-hosted security operations center that routes correlated security incidents to LLM agents, records every prompt, tool call and state …
Andrej Karpathy launched HackerTube at hackertube.org on October 31, 2026, a video site built from 6,454 videos recommended in Hacker News comments and sorted into 150 topics. Karpathy used Anthropic's Claude Opus 5.5 to…
A developer ran Captum's attribution methods on Cloud TPU v6e-1 hardware via TorchTPU and found that all captum.attr methods, including the LLMAttribution wrappers, execute on torch.device("tpu"), with the 27 determinist…
A developer published a guide on building a production-ready Retrieval-Augmented Generation (RAG) chatbot in Python using LangChain and OpenAI, covering the four-layer architecture of document ingestion, embedding genera…
Claude Code's background-agent summary feature sends roughly 40,000 input tokens to generate about 130 output tokens every 30 seconds, according to OpenTelemetry trace tests by observability platform Better Stack. The ag…
Meta's FAIR research team and New York University reported that their h1 system, a minimal agentic loop connecting frontier large language models to Lean 4 with real-time verifier feedback, scored 670 out of 672 on Putna…