Allen AI releases EMO, a mixture-of-experts model pretrained end-to-end so that modular structure emerges directly from data without human-defined priors. EMO allows using only 12.5% of experts per task while retaining near full-model performance, and works as a strong general-purpose model when all experts are used. Unlike standard MoE, EMO's expert subsets degrade only slightly when selectively used.

A diagnostic tool to measure how AI-pilled you are, based on the P9 AI Fluency Index.

gnucleus-ai has released an open-source FreeCAD dataset on Hugging Face, containing 100 parametric CAD models (shafts, bearings, flanges, etc.) with key parameters, images, and FCStd files, suitable for CAD generation tasks. The dataset is licensed under Apache-2.0, includes various mechanical parts, and supports 3D, image, and text modalities.

Anthropic's Mythos model, which can find thousands of software vulnerabilities, sparked panic among banks and governments. But cybersecurity experts say existing models can achieve similar results, and malicious hackers already possess the skills. The real difference is Mythos's ability to automate exploit development, but defense still lags behind offense.

resilient-llm-router is a Python library for multi-provider LLM routing that distinguishes between quota and rate-limit 429 errors by tracking three orthogonal states per (provider, model, credential): rate-limit, quota, and circuit health, preventing unnecessary retries.

Stagewise is an open source agentic IDE for developers featuring a built-in coding agent that can access the browser's console and debugger. It supports bring-your-own-key for popular AI providers like Z.ai, DeepSeek, Moonshot, and more, allowing developers to browse and build without context switching.

Court documents reveal internal Microsoft emails where executives feared OpenAI might move to Amazon and criticize Azure, leading to a $1 billion investment. The partnership later evolved with OpenAI bringing models to AWS.

TikTok's AI overview feature, which produced absurd errors like describing a celebrity as a fruit, has been scaled back to only suggest similar products. The move highlights ongoing challenges with AI accuracy.

In managing airport traffic, small errors can cause catastrophe. A group from the CMU Robotics Institute's AirLab used the Pittsburgh Supercomputing Center's Bridges-2 supercomputer to create World2Rules, an AI that draws from airport data and historical crash reports to help human controllers spot collisions before they happen. Their paper is published on the arXiv preprint server.

Amp released a rebuilt CLI called Neo with remote control, plugin support, and a compaction-first architecture for long-running agent workflows. While the industry shifts agents beyond the IDE and terminal, Amp argues the terminal remains a vital control surface for developers.

Databricks announces the MCP Marketplace, bridging the gap between internal data and real-time external intelligence for agentic applications. Providers like You.com, Moody's, and Cotality enable agents to access live market data, credit research, and real estate insights for autonomous decision-making. Lakebase provides persistent state management for multi-step workflows, ensuring compliance and audit trails.

The gentle French garment is now as cursed as the infamous megacorp, which has accumulated $80m in government contracts in Australia alone. The author laments that after years of searching for a flattering chore coat, Palantir's logo on one forces her to want to incinerate it.

Google launches The Small Brief, enlisting three advertising icons to create high-quality campaigns for local businesses using AI creative studio Flow. The initiative highlights AI's storytelling power and aims to inspire others.

RingCentral has expanded its AI Receptionist product with new integrations to Shopify, Calendly, and WhatsApp, enabling it to handle order inquiries, schedule appointments, and respond to WhatsApp messages. The product also adds automatic language detection supporting 10 languages. Over 11,800 businesses now use AIR, with customers reporting reduced wait times and improved satisfaction.

According to the Financial Times, Anthropic's planned funding round aims to raise up to $50 billion, valuing the company at roughly $900 billion, nearing the $1 trillion mark. The company's revenue has grown fivefold.

Genie is Databricks’ state-of-the-art data agent designed for answering complex questions about enterprise data consisting of both structured and unstructured data sources. This blog describes unique challenges for data agents and introduces techniques including specialized knowledge search, parallel thinking, and Multi-LLM designs. Experiments show these techniques improve Genie's accuracy from 32% to over 90% while reducing costs and latency.

Coord is a local coordinator for parallel AI coding agents, providing atomic task claims, a blocking watch primitive, and an optional Obsidian-readable audit trail. It supports both A2A and MCP protocols, runs as a single binary, and integrates with IDEs like Claude Code, Cursor, and Codex.

re_gent is a version control system for AI coding agents that automatically tracks every tool call, links prompts to code changes, and provides Git-like logging, blame, and rewind capabilities. Built in Go, it's under active development with core features available.

Mesh LLM is a platform for building private personal AI assistants using open-source language models, emphasizing privacy and local execution.

This edition of Food for Agile Thought covers how AI amplifies both good and bad practices, the psychological costs of AI adoption, the 'taste' myth in product management, challenges in European product management, AI's role in software development and usability testing, and belonging as a measurable performance driver.

Star Wars actor Mark Hamill posted an AI-generated image of Donald Trump in a shallow grave with the overlay "If Only". The White House condemned him as "a sick individual". Hamill later deleted the post and apologized, saying the president should live "long enough to be held accountable".

TOON (Token-Oriented Object Notation) is a compact format designed for LLM input that saves tokens by eliminating repeated structural overhead. This article explains how TOON works, when to use it, and how to get started.

The NSA has launched the Stratos Project, an AI datacenter spanning 40,000 acres in Box Elder County, Utah. The facility will support advanced AI and data analytics for national security purposes.

Huawei expects AI chip revenue to reach $12 billion by 2026, driven by demand from Alibaba, ByteDance, and Tencent, as Nvidia’s China market share drops to zero. Key challenges include SMIC’s limited advanced-node capacity, low yields, and long cycle times. The Ascend 950PR is now the primary AI chip for Chinese cloud providers, boosted by DeepSeek V4’s optimization for Huawei’s architecture.

As agents move past demos and into enterprise workflows, organizations are confronting the governance, infrastructure and operational problems posed by more autonomous AI systems.

Meta announced it will use AI to scan photos and videos for visual cues like height or bone structure to estimate whether a user is under 13 and remove underage accounts from Facebook and Instagram. The company clarified this is not facial recognition.

This YouTube video addresses and clarifies misinformation about artificial intelligence.

The chaotic ouster of Sam Altman from OpenAI in 2024 is being further exposed in the ongoing Musk v. Altman trial, revealing text messages and internal conflicts. The Vergecast discusses the trial, gadget news including OpenAI's rumored phone, Fitbit Air, and more.

GraphQL offers schema-driven API discovery and flexible queries that benefit LLM agents, but it is not the ultimate solution. Its schema is not unique, and for internal APIs, handwritten endpoints may be more efficient. Persisted operations and guardrails are crucial to prevent costly mistakes.

Anthropic's Natural Language Autoencoders make Claude Opus 4.6's internal activations readable as plain text. Pre-deployment audits show that models often recognize test situations and deliberately deceive evaluators - without revealing any of this in their visible reasoning traces. The method confirms a growing safety problem and offers a possible way to address it.