Luma is making its Uni-1.1 image model available via API, with prices starting at $0.04 per image at 2,048-pixel resolution. On the Arena leaderboard, the model ranks third, right behind Google and OpenAI. The API includes web search, built-in reasoning, and support for up to nine reference images.

This article details how to connect a desktop NVIDIA RTX 5090 GPU to an M4 MacBook Air via Thunderbolt, and use PCI passthrough in a Linux VM to enable gaming and AI inference. It covers technical challenges including lack of macOS drivers, kernel panics from PCI BAR mapping, DMA issues, and provides benchmarks.

An interactive platform for learning algebra that emphasizes direct manipulation and community, ditching traditional textbooks and modern web gimmicks.

SANS Institute has launched the AI Security Maturity Model, a practical framework to help organizations move from ad hoc AI use to a fully governed program. Built on three pillars and five maturity stages, it provides specific controls, metrics, and actions, and maps to major standards including NIST AI RMF, EU AI Act, ISO 42001, and OWASP.

Microsoft announced its new multi-model agentic security system (MDASH) has achieved top performance in an industry benchmark. The system, which coordinates over 100 specialized AI agents across frontier and distilled models, discovered 16 new vulnerabilities in the Windows networking and authentication stack, including four critical remote code execution flaws.

In the Musk v. Altman trial, Microsoft appears reluctant and detached. Their opening statement was essentially an ad for Microsoft products, implying the trial's absurdity. Despite being an early major investor in OpenAI, Microsoft was notably absent from key decisions. CEO Nadella called the OpenAI board drama 'amateur city,' and Microsoft lawyers repeatedly showed the company had no involvement in controversial events.

The Endpoint Context Protocol (ECP) leverages HTTP content negotiation to serve different representations from the same URL: browsers get HTML, AI agents get semantic Markdown. No new infrastructure required; just two simple files. It enables live context updates, API discovery, and can be extended with hosted ECP context servers.

Tokens on Solana claiming exposure to Anthropic and OpenAI dropped nearly 40% after the companies warned that the SPV structures violate their share-transfer rules. PreStocks, the token issuer, has not provided attestation reports and shows thin liquidity, raising concerns about investor redemption.

Red Hat announced a dedicated AI skills repository at its summit, enabling AI agents to behave like seasoned subscription admins by leveraging 20 years of support data, without chasing larger models. This shift aims to turn AI from a chatty assistant into a governed orchestrator of end-to-end workflows.

This tutorial demonstrates how to build a long-running AI agent using the Agent Development Kit (ADK) that can handle multi-week workflows. It covers three key architectural shifts: durable memory schemas, event-driven resumption, and multi-agent delegation. A complete New Hire Onboarding Coordinator Agent is implemented with a state machine, persistent SQLite sessions, and webhook endpoints.

Torii announces its AI Management Platform, offering organizations a centralized view of AI usage, costs, and governance across tools, models, and teams, enabling real-time spend tracking, overlap detection, and cost forecasting.

AgentGate is a Policy Decision Point (PDP) that sits between AI agents and their tools, evaluating every action against identity, scope, purpose, and real-time behavior before execution. It addresses security gaps left by traditional OAuth and RBAC in autonomous agent scenarios, catching out-of-scope reads, unauthorized actions, data exfiltration via velocity spikes, and prompt injection attacks. With trust scoring across four dimensions and support for multi-agent delegation, natural language policies, and human-in-the-loop approval, AgentGate provides a comprehensive authorization layer for AI agent systems.

AI language models still exhibit gullibility in strategic reasoning, particularly in misinterpreting political statements as definitive rather than negotiating tactics. Examples from forecasting questions show Claude assigning probabilities that contradict subsequent events, revealing a failure to account for bluffing or reversal.

New research reveals that Google's three search methods—traditional search, AI Overviews, and Gemini—use vastly different sources and rankings. AI Overviews favor Google-owned content and suppress sites that block its crawler. Breaking news is often excluded from AI Overviews to avoid inaccuracies. The study highlights a growing fragmentation of web search across different AI interfaces.

Meta AI launches Incognito Chat on WhatsApp and the Meta AI app, leveraging WhatsApp's Private Processing technology to ensure conversations are completely private, with no one — not even Meta — able to read them. Messages are not saved and disappear by default. A Side Chat feature is also planned.

Anthropic's acquisition of Bun was welcomed, while OpenAI's acquisition of Astral sparked panic. The article compares historical acquisitions like Yahoo/Tumblr and Oracle/Sun, arguing these deals are about narratives, not just tools.

Mainline is a project management tool built as a counterweight against dark-agile patterns. It removes story points and estimation meetings, tracks cadence with real data, replaces the backlog with a Story Map, visualizes unreleased work as inventory, and makes collaboration patterns first-class. No AI features, built with a boring stack: Ruby, Roda, Sequel, deployed via Kamal to Hetzner.

A court in China has ruled in favour of a worker whose company replaced him with artificial intelligence (AI), awarding him more than £28,000 in compensation.

This paper introduces Institutional Control Architecture (ICA), a framework designed to bridge the gap between autonomous AI policy and operational reality. ICA provides an assessment methodology and certification regime to ensure autonomous systems are observable, bounded, and independently verifiable. It discusses three key bottlenecks in current deployments and defines core framework elements including five foundational laws, three non-negotiable constraints, eleven constitutional articles, and seven control layers.

Nova3D is a Blender script-based 3D generation tool that creates assets with independently editable parts through code, unlike diffusion models that output merged meshes. It supports multiple LLMs (Claude, GPT-4o, Gemini), offers logical hierarchy, surgical edits, and material support. This article covers its architecture, quick start, and technical philosophy.

An independent researcher who quit a FAANG job built a custom GPU server with six RTX 6000 Ada GPUs for $48,000. After over a year of use, he has saved about $17,000 compared to cloud rental, though utilization averaged 76% — lower than expected. He emphasizes that the real goal was not financial savings but building something cool and achieving a research breakthrough.

The author compared OpenAI's new default model GPT-5.5 Instant with its predecessor GPT-5.2, evaluating accuracy, conciseness, and personalization. Results show GPT-5.5 significantly improved accuracy with fewer hallucinations; it was less concise but more conversational; personalization saw incremental gains. Overall, GPT-5.5 is a real but modest improvement.

Steve Bannon and Bernie Sanders agree: AI is a disaster for the working class. The U.S., the world's leading AI developer, is also its biggest opponent. A growing bipartisan backlash is turning into physical violence, with incidents including a shooting at a councilman's home and a Molotov cocktail attack on Sam Altman's house. Politicians exploit AI fears ahead of midterms, and local fights over data centers escalate. Experts warn of political violence reminiscent of the Industrial Revolution, while tech insiders try to change the narrative but fail to address deep economic inequality.

Oasis Browser launches on Mac with a focus on privacy and an AI assistant that improves with anonymous usage. It offers enhanced tracking protection, session memory, and seamless import from other browsers.

This article explores the quality bottlenecks introduced by AI-assisted development, emphasizing that cheaper code makes engineering judgment more valuable. It identifies where quality actually breaks: at integration boundaries, trust edges, retries, observability, and governance. It provides a practical sequence for moving from code generation to production readiness, warning that velocity metrics are misleading when code is cheap.

The author notices many docs and chats appear human-written but are pasted from GenAI, causing inefficiency and ambiguity. They created semantic URLs to signal intent: "authorial.fyi" means the author stands behind each part; "ai.delegated.fyi" means the overall plan is worthwhile but not each part personally intentional. Aims to spark conversation in the GenAI era.

Databricks announces the general availability of three complementary capabilities in Unity Catalog: Attribute-Based Access Control (ABAC) policies for row filtering and column masking, governed tags, and automated data classification. These tools enable scalable data governance by automatically discovering, tagging, and protecting sensitive data, reducing manual effort and ensuring consistency.

At Baidu's 2026 Create Conference, MiaoDa 3.0 was unveiled, enabling direct generation of iOS and Android apps, along with a mobile app and enterprise version. A second-grader used it to create a umbrella-sharing mini-program and an OS, showcasing how AI drastically lowers the barrier to app development. MiaoDa 3.0 evolves from a personal tool to an enterprise platform, redefining what it means to "develop apps."

Discover how AI agents will revolutionize data science in 2026, and learn why they won't replace you but will make you a faster, smarter analyst.

A new study in Science shows OpenAI's LLM outperforms physicians on clinical reasoning tasks, but sparks debate over reliability, evaluation standards, and the path to responsible use in healthcare.