AI brief — 20 May 2026

Andrej Karpathy, OpenAI founding member, starts at Anthropic this week

Andrej Karpathy — one of OpenAI's eleven founding researchers and the former head of Tesla's Autopilot programme — joined Anthropic on 19 May to work on its pre-training team under team lead Nick Joseph. He will build a new group dedicated to using Claude itself to speed up pre-training research — the most compute-intensive phase of building a frontier AI model — a signal that Anthropic believes AI-assisted research is the key competitive lever against OpenAI and Google. Karpathy left OpenAI in early 2024 to found Eureka Labs, an AI-native education startup, before joining Anthropic; CNBC reports he started this week.

Anthropic acquires Stainless, the SDK platform behind every official Claude API library

On 18 May, Anthropic announced it had acquired Stainless — the developer tools company that generates SDKs, command-line tools, and Model Context Protocol (MCP) servers used by hundreds of API providers, including OpenAI, Google, and Cloudflare. The Information reports the deal exceeded $300 million. Stainless's hosted SDK generation service will be wound down, though existing customers retain full rights to modify and use the SDKs they have generated. Anthropic said the acquisition aims to strengthen its control over the infrastructure that connects AI agents to external systems, as MCP adoption accelerates across the software industry.

Google and Blackstone form a $5 billion AI cloud company built on Google's own chips

Google and Blackstone announced on 19 May they are creating a new AI cloud computing company, with Blackstone committing $5 billion in equity — rising to $25 billion including leverage — for a majority ownership stake. The venture will offer cloud services, data centre capacity, networking, and Google's proprietary Tensor Processing Units (TPUs) as a compute-as-a-service product, positioning itself as a rival to neocloud providers such as CoreWeave. The company plans to bring 500 megawatts of computing capacity online by 2027. Blackstone's press release describes this as one of the largest single AI infrastructure investments of 2026.

arXiv will ban authors for a year if they submit unchecked LLM output

Thomas Dietterich, chair of the Computer Science section at arXiv, announced on 17 May that the preprint server will impose a one-year submission ban on any author whose paper includes incontrovertible evidence of unchecked AI output — such as hallucinated citations, AI meta-comments accidentally left in, or placeholder data marked "fill in with real numbers." Re-entry requires first clearing peer review at a recognised venue. Nature reports that the rate of hallucinated citations in arXiv preprints has risen roughly tenfold in three years, from one in 2,828 papers in 2023 to one in 277 by early 2026. The policy reflects growing institutional concern about academic integrity as AI writing tools become standard in research workflows.

Colorado replaces its landmark AI law with a narrower, industry-lighter successor

Governor Jared Polis signed Senate Bill 26-189 on 14 May — substantially revising Colorado's 2024 Artificial Intelligence Act, the first comprehensive AI statute in the United States, after two years of legal pressure including a federal lawsuit supported by the US Department of Justice. The new law narrows its scope from "high-risk AI systems" to "automated decision-making technology," covering software that materially influences decisions in hiring, lending, housing, healthcare, and essential government services; it drops the requirement to explain how AI reaches decisions, replacing it with a consumer notification and appeals obligation. The law takes effect 1 January 2027; its passage reflects the broader tension between state-level consumer protection and federal pressure for a lighter regulatory touch. (Follow-up on a two-year developing story; signed 14 May.)

Microsoft's AI chief predicts white-collar work is 18 months from full automation

Mustafa Suleyman, CEO of Microsoft AI, told the Financial Times this week that AI will reach human-level performance on most professional tasks within 12 to 18 months, naming accountants, lawyers, marketers, and project managers as particularly at risk and noting that software engineers are already delegating most code-writing to AI. The claim contrasts with a recent study by AI evaluation non-profit METR, which found that AI-assisted software developers took 20 per cent longer on tasks than those working without AI — suggesting that productivity gains in knowledge work remain uneven and context-dependent. Futurism notes that while AI coding tools are now standard, measurable automation of professional knowledge work at scale has yet to materialise.