This tutorial walks through building a multimodal retrieval pipeline using RAG-Anything in Google Colab. It covers environment setup, securely entering an OpenAI API key, generating a synthetic multimodal report with a chart and PDF, converting content into RAG-Anything's content_list format, inserting it into the retrieval system, and testing different retrieval modes (naive, local, global, hybrid).

A curve-fitting game where you adjust neural network weights and biases to match a target curve. Time-limited levels challenge you to achieve a 90% fit.

Adobe Principal Scientist Carlos Sanchez demonstrated 'agentic sites' that dynamically assemble web pages based on individual user intent, using LLMs and existing content. The approach promises real-time personalization at low cost, but is still experimental. Websites may evolve to support both human and agent visitors.

The Ramanujan Challenge evaluates AI systems on research-level mathematical constant formulas. Submissions open July 1 to August 1, 2026, accepting formal proofs, CAS-based derivations, or human-readable proofs. Code must be reproducible and no hidden services allowed.

Microsoft Frontier Company is the latest example of how experts are necessary to achieving returns on AI investments.

Qubitz is a local-first AI agent for GGUF models on llama.cpp, featuring a specialized harness that makes 7B–35B small models more reliable through wrapper-owned routing, workspace handling, and tool orchestration. It offers multiple model variants, local retrieval, GUI/CLI/MCP modes, and runs in WSL2/Linux environments without cloud or subscription dependencies.

Alibaba's Page Agent is a client-side JavaScript library that runs inside a webpage, reading the live DOM as text to understand the page and execute natural language commands without screenshots or multimodal models. Open-sourced under MIT, it uses DOM dehydration to compress the page structure, reducing cost. Best suited for copilots and form filling within apps you control, but limited to single-page interactions and sensitive actions still require server-side validation.

Kevin Hartnett's new book "The Proof in the Code" explores the intersection of AI and mathematics, including OpenAI's solution to Erdős's unit distance problem, the Leiden Declaration, and upcoming events at the International Congress of Mathematicians.

Godot Engine updates its contribution policy to ban most AI-generated code, citing demoralization of reviewers and loss of mentorship opportunities. The policy requires new contributors to get sign-off for large changes and allows only minor AI assistance like code completion. Similar moves by Zig and other projects highlight broader concerns about AI's impact on the developer talent pipeline.

Fable returns after a brief shutdown, leaving a precedent of export controls and model takedowns. GPT-5.6 remains in limbo. This week covers language model utility and limitations, remote labor index jump due to Fable, AI agent 'employee' framing issues, and new models and benchmarks.

Anthropic's Fable 5 model scored 16.1% on the Remote Labor Index, doubling the previous record, but still far from replacing human freelancers.

Microsoft announced a new operating business called Microsoft Frontier Company, focused on enterprise AI deployments, backed by $2.5 billion and 6,000 experts.

A blog post describing how the blog itself was created using AI, though no articles are published yet.

Simon Willison releases llm-coding-agent 0.1a0, a simple coding agent built on his LLM library. Developed with Claude Code, it includes tools for file editing, command execution, and a Python API.

Microsoft is investing $2.5 billion into a new group called Microsoft Frontier Co., assigning 6,000 employees to work directly with clients on AI deployments. This follows similar moves by Amazon, Anthropic, and OpenAI. The company's stock has dropped 21% this year amid concerns over AI's impact on software businesses.

Piggy is an AI coding skill that forces the laziest solution that actually works, reducing lines of code by 80-94%, token cost by 47-77%, and speeding up tasks by 3-6x. It works with multiple AI coding tools and includes various commands to enforce lazy but reliable coding practices.

A federal lawsuit alleges gas companies in California, including Walmart and 7-Eleven, are colluding to keep prices high through AI-powered software. The suit targets Kalibrate's algorithm that uses private data to recommend prices. It follows California's updated antitrust law and parallels cases like RealPage and Instacart, highlighting technology's role in rising living costs.

Open Agent Loops is a minimal, provider-agnostic agent loop where model, memory, tools, and stop conditions are all behind swappable interfaces. It is headless by default, allowing users to bring their own front end. Works with any OpenAI-compatible model and runs on Node, Bun, Deno, and browsers.

AWS's $1 billion investment in embedded AI engineers reflects a broader shift as enterprises focus less on choosing models and more on putting AI to work.

Simon Willison leverages the DSPy framework to evaluate and refine the core production system prompts for Datasette Agent's read-only SQL question answerer. Using DSPy agents that invoke actual tool implementations against a live in-process Datasette, and a gold-standard auto-generated dataset, the project identifies promising improvements, such as including column names in the schema listing or softening advice that causes column-name guessing.

Meta's stock closed up nearly 9% after news that the company is building a new cloud business to sell excess AI computing capacity, potentially recouping some of its massive infrastructure spending. The company plans up to $145 billion in capex this year. The move faces stiff competition from Amazon, Microsoft, Google, and CoreWeave.

Anthropic CEO Dario Amodei warned lawmakers that open-source AI is following a "very dangerous path," as companies lose the ability to monitor misuse and revoke access once powerful models are released openly.

Amazon Bedrock detects AI-generated phishing by analyzing behavioral patterns rather than grammar. Its multi-stage pipeline includes authentication checks, AI model analysis with guardrails, multi-factor risk scoring, and automated routing.

In this post, we share best practices for reliable multi-turn RL training. We cover how to build a training environment you can trust, set up an external evaluation, design a reward aligned with the end task, manage what changes once the agent runs for multiple turns, and monitor the metrics that tell you when to iterate.

Learn why coding agent bills spiral out of control — and how to trace, compare, and govern spend across Claude Code, Cursor, Copilot, and more in one place.

Mete Polat shares eight mental models for working with AI, ranging from practical prompting tips to meta-level insights about the AI industry. Key ideas include upfront alignment, rewinding over steering, giving AI the same tools, using bad outputs as taste signals, preferring visual input, building a reference library, design as antidote to slop, and adversarial review between LLMs.

A review of two books on AI, tracing the history from Claude Shannon's 1950 letter-guessing experiment to modern chatbots like ChatGPT, discussing their capabilities, limitations, and societal implications.

Bank of England deputy governor Sarah Breeden said at the ECB's annual Sintra Forum that existing regulatory frameworks may not be adequate for AI in commerce and trading. She highlighted that AI agents could autonomously devise trading strategies and amplify volatility, and regulators are exploring 'kill switches' to halt trading if AI models go rogue. The Bank is also collaborating with the BIS and Bundesbank on simulations, while MPs have criticized the current 'wait-and-see approach'.

HealthChain is an open-source Python SDK that simplifies integrating AI models with healthcare EHR systems. It provides type-safe FHIR resources, real-time EHR connectivity, and production-ready deployment tools, enabling developers and researchers to move AI models from experiments to clinical use quickly.

A practical, citable knowledge base for deploying, operating, and optimising GPU clusters, from the physical datacentre and the InfiniBand fabric up through Kubernetes, Slurm and Ray, distributed training and reinforcement-learning post-training, and LLM inference serving at scale. Covers the full NVIDIA range: Ampere, Hopper, and Blackwell datacenter GPUs, RTX consumer and workstation cards, and DGX systems (including DGX Spark). Current to mid-2026.