IEEE Spectrum editor Harry Goldstein recounts his more than 20-year collaboration and friendship with Robert N. Charette, the risk ecologist whose 2005 article “Why Software Fails” became a classic. Charette made software visible through his Risk Factor blog and deeply reported features, and now retires to nature photography and fiction, guided by his maxim: assumptions made are risks accepted.

In this LessWrong post, Anthony DiGiovanni argues that the term 'equilibrium' is often used ambiguously in AI bargaining research, and suggests 'tabooing' it to encourage clearer conceptual framing.

A New York school district has paused plans to bring an android into the classroom. Dave Schilling’s satirical take asks whether robot teachers are really the future—and argues that a child’s wellbeing is too important to hand to a machine.

AI's capability gap follows one rule: it becomes superhuman wherever verifying an answer is cheap and mediocre wherever verification is expensive. With web text nearly exhausted, synthetic data and RL on verifiable outcomes amplify the asymmetry, while slow domains lose the senior judgment they depend on.

Vector databases store embeddings and power similarity search for AI applications. This guide explains the fundamentals of vector databases and similarity search, then covers LanceDB's key features—multimodal storage, multiple index types, hybrid search, versioning, schema evolution, and object storage—with Python examples for vector search, PDF ingestion, and image retrieval.

Weightlift is a small library that streamlines model weight downloading, caching, and progress display for on-device web-based AI apps, supporting transformers.js, ONNX, prakeet, and more.

Supabase has open sourced supabase/evals, an Apache-2.0 benchmark that runs coding agents such as Claude Code, Codex, and OpenCode on real Supabase tasks in containerized stacks and scores them with deterministic checks plus LLM-as-a-judge. It powers a public leaderboard and daily regression suite.

SendPage turns AI-generated HTML documents into shareable links that open in one tap inside WhatsApp, WeChat, or email. It supports PDF, image, and Word export, lets you edit text after sending without changing the link, and offers a free tier of 5 documents per month.

MiniMax has launched MiniMax H3, a general-purpose omni-modal model that generates 4–15 second 2K videos with native stereo audio from text, images, video, and audio references. It is available via API and the Hailuo AI app, with open weights promised but not yet shipped. Early benchmarks rank it first in video editing.

Chinese AI researchers are increasingly joining X to share technical insights, build international profiles, and connect with the global AI community. The article explores why Chinese platforms fall short, how this trend fills a gap left by OpenAI and Anthropic researchers, and the branding benefits for Chinese AI labs.

Discussion | Link

Discussion | Link

Discussion | Link

Thomson Reuters unveiled Thomson, its own AI model, reporting benchmark results that put it on par with the strongest frontier models, including Claude Opus 4.8, and ahead of GPT-5.5, Claude Sonnet 5, and Gemini 3.1 Pro. Built on an open-source foundation and trained on proprietary Thomson Reuters content, Thomson will first appear in CoCounsel Legal's Tabular Analysis in August, with broader integration planned across legal and tax products.

ACE sidecar is a local reverse proxy that measures and optimizes token usage and cost for AI coding agents like Claude Code. It leaves traffic unchanged, uploads nothing, reads existing transcripts on disk, and provides a phased optimization roadmap grounded in real session data.

Discussion | Link

US House committees have asked DoorDash to disclose its use of Chinese AI models and related security testing by August 14, after a co-founder said Moonshot AI's Kimi K2.6 outperformed US models in internal code review. Lawmakers acknowledge practical benefits but warn of national security risks, with DoorDash now the third US company under scrutiny.

Leopold Aschenbrenner's AI hedge fund Situational Awareness collapsed from $45 billion to about $10 billion in assets driven by falling semiconductor stocks and margin calls, forcing the sale of leveraged bets to Citadel at a discount.

Snapchat said it will stop recommending wholly AI-generated videos in its Spotlight feed in favor of human-made content, joining YouTube, LinkedIn and Substack in a growing effort to combat low-quality AI-generated content. The platform still allows AI-enhanced or edited content to appear in recommendations.

Urldn is a URL shortener with built-in analytics and an AI data analyst, designed to help users track what actually drives clicks. It is currently open via a waitlist.

The AI Visibility Evidence Model is a publisher-side reference model that orders factors affecting AI visibility by strength of evidence. It defines five factors—Topical Relevance, Machine Access, Entity Consistency, Extractability, and Independent Corroboration—each assigned a documented evidence grade based on peer-reviewed research, controlled preprints, and official platform documentation. The model is descriptive, not prescriptive, and emphasizes priority rather than ranking.

The author observes xAI is building a computer-use tool similar to OpenAI's ChatGPT feature, but argues the real next step is proactive AI that acts before being asked. He introduces Orb, an open-source proactive AI platform. The iOS app is live, connects to your PC, and uses Claude Code, Codex, or Grok Build to act on your behalf. It scans activity every 15 seconds, runs an LLM call every 12 hours, and learns from your feedback to deliver useful notifications. The author invites feedback and contributions.

Kimi K3 is Moonshot AI's 2.8T-parameter open-weight model—the first in the 3-trillion-parameter class. Here's how it's architected, what it costs on Together AI, and how to call it with code examples for reasoning effort, streaming, vision, structured output, tools, and dynamic tool loading.

OpenAI shares new results on long-standing open problems in mathematics and theoretical computer science, including advances in geometry, cryptography, and complexity.

IFPI, the global recording industry body, has introduced eligibility principles that require AI-assisted recordings to be lawful, substantially human-made, and free of stream manipulation to enter official charts. The rules first apply to IFPI-run charts across Latin America, the Middle East, Africa, and Southeast Asia, with expansion to more than 20 national chart programs.

The new stateless MCP 2.0 specification dramatically simplifies client and server implementations. Simon Willison built mcp-explorer, datasette-mcp, and llm-mcp-client around it, and argues MCP's auditable, constrained tools are a safer alternative to giving agents arbitrary shell and curl access.

With AI coding agents, you can build a complete project first and learn full stack development by tweaking and extending it. The article uses bestbooks.guide as an example and combines observing, writing, and doing, along with roadmap.sh and spec-driven development, to master React, Node.js, PostgreSQL, AWS, and Terraform.

Google has paused its new AI tool that lets people create fake images on Google Earth, less than 48 hours after launch, following a BBC Verify report in which experts warned it could be misused to spread misinformation.

Amazon Quick announces the Agentic Catalog Experience, an AI-powered workflow for data curators to discover upstream catalog assets in natural language and auto-create Datasets and Topics with inherited semantics. Now in preview for AWS Glue Data Catalog and Databricks Unity Catalog.

SQLite.ai offers two independently purchasable products: SQLite Cloud, a fully managed SQLite database platform, and CloudSync, an offline-first sync service that works with SQLite Cloud, PostgreSQL, or Supabase. This page details their plans, resources, and pricing.