Spice is an open-source decision layer above AI agents that controls actions before execution. It observes context, detects conflicts, simulates options, and dispatches tasks to the right agents through a loop: perception → state model → simulation → decision → execution → reflection.
Simon Willison investigates Monty, a sandboxed Python subset implemented in Rust. Using Claude Code to analyze the latest release, he confirms that resource limitation parameters (max_duration_secs, max_memory, max_allocations, max_recursion_depth) work as advertised.
Lawyers continue citing fake court cases generated by AI models in legal filings. The problem persists despite warnings, exposing LLM limitations in producing factual content for high-stakes applications.
A developer tests the minimum model size for orchestrating a local ReAct loop. Qwen3.6-35B-A3B (MoE, ~3B active) is his threshold: below it, the model invents tool parameters or overgeneralizes calls. He improves accuracy by exposing exact signatures in the system prompt.
OpenAI launches Appshots, a macOS feature enabling users to send any app window's contents to Codex with a single click. Codex receives the context needed to complete coding tasks.
Experimental llama.cpp fork optimizing MoE for 12GB VRAM GPUs. Author selectively loads experts to VRAM instead of full layers, reaching 26 tk/s on RTX 2060 (vs 19 tk/s default) with 62% hit rate. Seeking testers on 3060/4060.
llama.cpp supports asymmetric KV caches (q8/q4) but currently falls back to CPU processing instead of GPU with CUDA for certain combinations. User evaluation shows q8_0/q4_0 costs only 1.3% precision loss while reducing memory by over 50% vs f16/f16.
Spotify and Universal Music Group officially launch AI-generated remixes and covers on their music streaming platforms.
Microsoft and EY announce a $1 billion investment over 5 years to accelerate industrial AI deployment in enterprises.
Open-source tool converting code into interactive, explorable knowledge graphs. Compatible with Claude Code, Cursor, Copilot, Gemini CLI, and other editors.
Workmux combines git worktrees and tmux windows for frictionless parallel development. Open-source tool enabling simultaneous management of multiple work branches with native tmux integration.
MemOS is a self-evolving memory OS for LLMs and AI agents. Features ultra-persistent memory, hybrid retrieval, and cross-task skill reuse with 35.24% token savings.
Repomix is a tool that compresses an entire repository into a single file optimized for LLMs. Compatible with Claude, ChatGPT, DeepSeek, Perplexity, Gemini, and other AI models.
AWS Labs releases aidlc-workflows, a framework of adaptive steering rules for directing AI coding agents. The project provides AI-driven lifecycle workflows with rule-based steering to improve code agent quality and reliability.
France ranks 2nd in declared AI maturity in Europe, yet 49% of projects remain in pilot stage and 80% have no measured ROI. Gap reveals disconnect between stated adoption and actual impact.
Lemon-mlx-engine integrates ROCm 7.13 to run LLMs locally on AMD GPUs. Update includes bug fixes and kernel fixes for Qwen3, 3.5, and 3.6 MoE.
Three AI infrastructure startups reach unicorn status: Exa (vector search), Modal (cloud platform), and TurboPuffer (distributed cache). Major funding rounds confirm consolidation in the AI infrastructure market.
CODA rewrites transformer blocks as GEMM-Epilogue programs to optimize inference. The technique fuses matrix operations and post-processing into a single GPU primitive, reducing latency and memory bandwidth.
Simulation study on double descent phenomenon in linear regression with contaminated data. Authors compare least-squares interpolation (non-robust) against robust alternatives. Finding: overparametrization enables double descent, least-squares estimator outperforms robust methods despite outliers.
Empirical study (20 users) of a multimodal conversational agent with emotion recognition. Computer vision and linguistic analysis integrated. Key finding: text analysis outperforms facial recognition (poker face effect). Miscalibrated proactivity reduces user engagement.
OpenAI named leader in Gartner's 2026 Magic Quadrant for Enterprise AI Coding Agents. Codex recognized for innovation and enterprise-scale deployment.
Google API keys remain active for up to 23 minutes after deletion. This propagation delay creates a vulnerability window where deleted keys can still be exploited.
A technique to extract preferences from a repository's merge history. Assuming accepted revisions incrementally improve code quality, preference signals can be distilled to align AI agents with institutional practices, avoiding expensive expert annotation.
Google DeepMind launches an Accelerator program in Asia Pacific to develop AI solutions addressing environmental risks. The initiative provides funding, technical expertise, and Google Cloud access to local startups and researchers.
Production liveness detection systems rely on outdated threat models (static images, basic replays). Current synthetic media quality far exceeds historical training data. Critical question: can models trained on legacy deepfakes generalize to generation techniques that didn't exist when training data was collected?
Google tests how websites handle AI agents through a new experimental 'Agentic Browsing' category in its Lighthouse analysis tool. The test includes checking for llms.txt file presence.
LatitudeGames releases Equinox-31B, a Gemma 31B fine-tune optimized for interactive storytelling. The model blends dark adventure and slice-of-life narrative data. Available on Hugging Face in GGUF format, accessible via aidungeon.com with subscription.
Chrome DevTools MCP: a Model Context Protocol enabling AI agents to interact directly with Chrome DevTools for real-time debugging and inspection of web applications.
Release of datasette-agent-charts 0.1a2. Adds "View SQL query" buttons below rendered charts to inspect generated SQL queries.
Release of datasette-agent 0.1a3, an extensible AI assistant for Datasette. Alpha version enabling AI-powered interaction with databases through agents.
SAP partners with Mistral AI to simplify customer migration to S/4HANA. Mistral AI models help streamline the legacy software migration process.
Heretic Free Software Project received a legal notice from Meta regarding Llama model derivatives. The project removed model weights from controlled repositories and is diversifying infrastructure with mirrors on Codeberg and other platforms to preserve access independently of service providers.
Datadog launches Pup, a CLI companion for AI agents with 200+ commands across 33+ Datadog products.
Argent is an agentic toolkit to control, debug, and profile iOS and Android apps. Built by Software Mansion.
Stitch-Skills is a library of Agent Skills designed for the Stitch MCP server. Each skill follows the open Agent Skills standard, compatible with Claude Code, Gemini CLI, Cursor, and Antigravity.
Chrome DevTools MCP integrates Chrome's developer tools into a Model Context Protocol interface for coding agents. Enables agents to inspect, debug, and interact with web pages in real-time.
Unofficial Python API for Google NotebookLM providing full programmatic access to features, including those not exposed in web UI. Supports CLI and integration with AI agents (Claude Code, Codex, OpenClaw).
Vercel adds anomaly alert access via CLI with `vercel alerts` command. The `--ai` option displays AI investigation results for each alert. Available on Observability Plus.
AMD launches Ryzen AI Halo Developer Platform and Ryzen AI Max PRO 400 Series processors for next-generation agent computers. Official announcement detailing availability of Halo Box and AI 400 series.
User praises Qwen3.6 27B quantized Q5_K_XL on llama.cpp with dual RX 9070 XT GPUs. Model excels at debugging complex code (distributed backend services), achieving 398 tokens/s prompt eval and 46.9 tokens/s generation. Strong agentic capabilities despite low quantization.