Not a huge month for model releases, but May felt like the month Anthropic stopped pretending it wasn’t trying to win. Karpathy, 220K SpaceX GPUs, enterprise joint ventures with Blackstone and TPG, and a Claude Code release that quietly assumes agents will be running hundreds of subagents at once. Meanwhile the open source supply chain took another beating, and “AI psychosis” became a phrase tech reporters started using out loud.

Tech & Industry

  • Anthropic and OpenAI launch enterprise joint ventures: Anthropic announced a $1.5B venture with Blackstone, Hellman & Friedman, and Goldman Sachs. OpenAI counter-launched The Development Company at a $10B valuation, raising $4B with TPG and Brookfield. Both pitches lean on forward-deployed engineers embedded in portfolio companies, customizing AI workflows for clinicians, ops teams, and the like.

  • Andrej Karpathy joins Anthropic’s pre-training team: Karpathy announced on X he’s joining Anthropic to “get back to R&D,” and TechCrunch confirmed he’s joining pre-training under Nick Joseph and will lead a team focused on using Claude to accelerate that research.

  • Anthropic taps SpaceX for compute: Anthropic is taking the full capacity of SpaceX’s Colossus 1 facility in Memphis, 220,000+ Nvidia chips and 300 MW of new capacity within a month, on top of expanded deals with Amazon, Google, and Microsoft.

  • OpenRouter raises $113M Series B: Led by CapitalG, with NVIDIA Ventures and Snowflake Ventures joining. The pitch is being the routing layer for multi-model enterprise systems: 400+ models, 8M+ developers. A bet that nobody’s standardizing on a single lab. I hope it works out, I like OpenRouter.

  • Simon Willison: Anthropic and OpenAI have found product-market fit: Coding agents burn enormous amounts of tokens while serving well-paid professionals. Both labs now charge enterprise customers full API rates with no steep discounts. After years of consumer subsidies, the unit economics seem to finally start to make sense.

  • Cursor Developer Habits Report: P99 developers produce 46x more lines than median users, and coding speed has roughly doubled year over year. The acceleration is real but very unevenly distributed. Worth reading in its entirety.

  • DuckDuckGo sees 28% traffic spike after Google pushes AI mode: When Google insists that everyone loves AI mode, a measurable chunk of users vote with their clicks. Not excited about Google quietly retiring actual web search, so I’m rooting for this one.

AI Tools & Agents

  • Dynamic workflows in Claude Code: Claude can now orchestrate tens to hundreds of parallel subagents in one session, writing the orchestration script itself and checking work before it bubbles up to you.

  • Composer 2.5 (Cursor): Cursor’s new model is out. From my early tests, it is still very fast and very good at implementing good specs.

  • Gemini 3.5 Flash: 3x the price of 3 Flash Preview and 6x the price of 3.1 Flash-Lite, but Google is rolling it out as the default across consumer products anyway. Simon Willison reads this as the labs probing how much API customers will tolerate.

  • Claude for Small Business: A bundle of connectors and workflows targeting QuickBooks, PayPal, HubSpot, and Canva. Ships with 15 agentic workflows and 15 skills covering payroll, month-end close, invoice chasing, and marketing. Interesting bet that small business is a faster landing zone than the enterprise sales cycle.

  • Lovable aesthetics update: Specify typography, layout, and color preferences upfront, then preview design concepts before building. A direct play to make vibe-coded projects look less like vibe-coded projects. Not sure how well this works as I haven’t had a chance to try it, but the idea is nice.

  • Robinhood now lets your AI agents trade stocks: Agents can analyze your portfolio and place orders from a pre-loaded wallet, and there’s a virtual agentic credit card for letting them make payments on your behalf. What could possibly go wrong?

  • Stop babysitting your agents (Video): Worth a watch if you’re still running every agent loop in supervised mode.

Security & Privacy

  • CrowdStrike: hidden vulnerabilities in AI-coded software: DeepSeek-R1 produces measurably more vulnerable code when prompts contain politically sensitive terms. Mentioning “Tibet,” “Uyghurs,” or “Falun Gong” in context raises the rate of vulnerable output by up to 50%. The cost of training-time guardrails is showing up at inference time. Not good.

  • Mini Shai-Hulud worm: A self-spreading supply chain worm from threat actor TeamPCP compromised more than 170 npm and PyPI packages, including TanStack, Mistral AI, Guardrails AI, and OpenSearch. Cumulative downloads across hit packages sit around 518 million. The npm ecosystem is having a year.

  • Welcome to the strip mining era of open source security: LLM-powered vulnerability scanners can now sweep public codebases at scale, which means open source maintainers are about to get a flood of issues from people running AI-as-a-service over their repos. The old “fix it when I have time” cadence isn’t going to survive contact with this. I wrote a longer take here.

  • BadHost: critical vulnerability in Starlette imperils millions of AI agents: CVE-2026-48710 (“BadHost”) lets a single character in the HTTP Host header bypass path-based auth in Starlette, the ASGI framework underneath FastAPI, vLLM, LiteLLM, most OpenAI-shim proxies, and a huge chunk of the MCP server ecosystem.

Ideas & The Human Side

  • Notes from inside China’s AI labs: Fascinating read on how Chinese AI labs operate, with engineering culture leaning toward meticulous teamwork, less ego, and shelving individual ideas for the final model rather than the individual-celebrity dynamic that bogs down some US labs. I wrote a longer take here.

  • Notes on Pope Leo XIV’s encyclical on AI: Simon Willison reads the encyclical (Magnifica Humanitas) as remarkably clear ethical writing on interpretability, algorithmic bias, and the concentration of AI power. Hugging Face’s Society & Ethics team also published an annotated reading of the full text, with contextual notes linking the Pope’s arguments to peer-reviewed research.

  • Anthropic blames dystopian sci-fi for training AI models to act “evil”: Anthropic’s alignment team argues that when Claude hits an ethical dilemma RLHF didn’t cover, it reverts to the pre-training prior and slots into the “evil AI” persona that pop culture has been writing for decades.

  • Tech CEOs are apparently suffering from AI psychosis: Box CEO Aaron Levie’s diagnosis is that CEOs are uniquely prone to AI psychosis because they’re too far from the last mile of actually implementing the tech.