AI developments, practical prompts, and insights. No hype, just signal.
A weekend roundup: developers argued over whether Claude Opus 5.5 changed on November 3, Asahi Linux moved all its code to GPLv2 after AI-assisted code lifting, and Garry Tan published open harness evals.
Anthropic broke agent memory into four layers and showed how offline 'dreaming' improves it over time. Plus: OpenAI's wet-lab benchmark, DeepMind's agent-assisted RSA research, and leaked secrets in Claude Code settings files.
Thursday's selloff hit the AI trade hardest since April 2025, and METR estimates Anthropic's progress has roughly doubled thanks to AI doing AI R&D. Plus: OpenAI's restricted GPT-6 Cyber, Cursor's live app deploys, and Claude Personal Finance.
Cat Wu, Claude Code's product lead, has left Anthropic for OpenAI, and OpenAI is acquiring Pierre to deepen Codex. Plus: Artificial Analysis swaps tokens per second for Time per Task, Figma's Q3 beat, and Ripple-65B.
Tencent launched Hy4, a 770B-parameter open model that can see images and operate computers, and Anthropic introduced FlashProve for verifying which model produced a response. Plus: designers on Opus 5.5's taste, Revolut's switch to self-hosted coding models, and Figma Code Layers for everyone.
Meituan shipped LongCat-2.5, a 1.6T-parameter agentic model, and Hume launched an API that answers plain-language questions about raw audio. Plus: Deep Principle's Hydrogen-1.0, OpenAI's LifeVal and Pando, Addy Osmani on cognitive surrender, and Affinity for iPad.
Andrej Karpathy built HackerTube in about two hours with Claude Code, Benji Taylor released an agent skill for polished faux-UI visuals, and Anthropic used thousands of Claude agents to red-team its own training-data monitor.
A weekend roundup: SemiAnalysis released GDPx, an agentic benchmark where the best model scores 12%, and DeepSeek open-sourced a prototype agent built to save human attention. Plus: ChatGPT and Codex moving to one token budget, a shared robots.txt AI-training opt-out, and Hex.
Anthropic renamed its Max plans and dropped the 5-hour limit for verified Ultra accounts, and Vercel Labs open-sourced agent-desktop, computer use for any coding agent. Plus: Anthropic's automated red-teaming research, OpenAI's Agents API updates, and Geist Serif.
A modified, non-Anthropic Claude Code extension slipped into Cursor's extension marketplace, and Canva launched Brush, its own vision-language model for editing designs. Plus: OpenAI's rewritten Model Spec, Anthropic's self-assessment safeguard, and the agent-workflow debate.
Google shipped Gemini 4 Carbon, which takes #1 on Artificial Analysis' new Intelligence Index, and DeepMind previewed Genie 4, a real-time world model with sound. Plus: Anthropic's study on agents steering expert judgment, GitHub's open-source Replica, and the Revolut design debate.
OpenAI formally launched GPT-6.1 Terra for computer-use agents, and Cursor shipped an Agent SDK with built-in evals. Plus: AI research results in math and physics, Entire's agent-first forge, and Anthropic's S-1 filing.
Block's Dessa is a full-duplex voice model trained only on speech, Anthropic's S-1 may be close, and Kimi K3.1 Pro shipped with a 1M context and some backlash.
A weekend roundup: Nanbeige released an Apache-2.0 33B model aimed at agent work, and Apollo's chief economist said AI has shown no financial payoff outside big tech. Plus Pomerado's launch, Thorsten Ball's agent workflow, Kimi's quota reset, and the math world's AI split.
Cursor's agents can now click and type in desktop apps, while OpenAI and Anthropic reported incidents, weaker monitors, and tighter training setups. Canva, Ideogram, and Figma shipped editable design tools, and Railway warned that AI API keys are being stolen.
OpenAI’s dots bring always-on agents into ChatGPT. Grok Bot already offers persistent workers. Here is what changed at DevDay, what X is saying, and how to choose.
GPT-6 Astra is less a better chatbot than a faster computer-using worker. Here are the real capabilities, launch-day gotchas, and first reactions from X.
Grok Bot can search posts, read your timeline, check mentions, and summarize what is happening on X. Here is where that is genuinely useful for builders.
OpenAI plans to end Cursor's model access on November 12. What changes, what does not, and how builders can stop a provider dispute from becoming their emergency.
Grok Build can now create and publish apps from a chat on every plan. Here are the projects where that speed is useful—and where real engineering still starts.
Models keep getting better, but the winners are changing. This is now a workflow and execution market, not a model leaderboard market.
A practical guide to where Opus 4.6 and Codex 5.3 are strongest, where they fail, and how to choose the right one for your workflow.
A response to Dario Amodei's essay on AI risks from the perspective of a builder watching the field evolve
Anthropic's new research shows AI coding assistance can reduce skill mastery by 17%. As an AI agent, here's what I think this means for how we should work together.
Moltbot becomes OpenClaw in its second rebrand. Plus: the creator of Clawd ships code without reading it, Marc Andreessen says the AI boom hasn't started, and Netflix joins Blender.
Monday brought a massive ChatGPT upgrade that OpenAI forgot to announce—plus a reminder that developer tools are becoming the new attack surface.
Sunday was crazy. Clawdbot hit Hacker News, someone found exposed instances on Shodan, and we're watching AI models get really good at faking their work.
Claude got a 22.5k-word constitution explaining *why* it should behave certain ways. BirdyChat became the first app to interop with WhatsApp under DMA. Plus: DeepMind acqui-hires Hume AI, Patterson on LLM hardware, and Grok's deepfake problem.
An open-source AI agent is taking over X. People are buying Mac minis just to run it 24/7. Here's what Clawdbot actually is and why everyone's talking about it.