AI Daily Brief — 21 May 2025
The largest acquisition in OpenAI’s history landed the same day Mistral set a new open-source coding SOTA and LMArena raised $100M. Google I/O Day 2 dropped Gemini Diffusion. Microsoft Build Day 3 shipped an autonomous coding agent and opened MCP plumbing across Azure and Windows.
Top stories
- OpenAI acquires Jony Ive’s io in a $6.5B all-equity deal. OpenAI’s largest acquisition ever. Pays ~$5B in new equity (had already owned a 23% stake from Q4 2024). ~55 hardware engineers, software developers and manufacturing experts join OpenAI; Ive and LoveFrom remain independent while taking on deep design responsibilities across OpenAI and io. via Bloomberg
- Sam Altman + Jony Ive release 9-minute teaser film. Released alongside the acquisition. Altman frames the goal as creating an AI-first hardware device ‘as revolutionary as the iPhone’ but that brings ‘peace and calm’ instead of dopamine-chasing notifications. No product, form factor or timing was disclosed; reports later confirm a screenless / voice-first ambient device targeting H2 2026. via TechCrunch
- Mistral + All Hands AI ship Devstral — open-weights coding SOTA. Apache 2.0 agentic coding LLM at 46.8% on SWE-Bench Verified — beating the prior open-source SOTA by >6 points. Runs on a single RTX 4090 or 32GB Mac. Research preview; a larger agentic coding model promised in coming weeks. via Mistral
- LMArena raises $100M seed at $600M, a16z + UC Investments lead. The UC Berkeley-affiliated crowdsourced AI benchmarking platform that OpenAI, Google and Anthropic rely on goes from academic project to private company. Lightspeed, Felicis and Kleiner Perkins also participated. via TechCrunch
- Google I/O Day 2: Gemini Diffusion text-diffusion at ~1,479 tok/s. Experimental research model that generates text by iteratively refining noise instead of token-by-token autoregression. ~5x Gemini 2.0 Flash-Lite; 89.6% HumanEval, 76.0% MBPP, 23.3% AIME 2025, 40.4% GPQA Diamond. Widely called the sleeper hit of I/O. via DeepMind
- Microsoft Build Day 3: GitHub Copilot SWE agent + NLWeb open protocol. Enterprise-grade asynchronous coding agent that takes assigned issues, runs autonomously in a secure container, and files draft PRs (requires human approval before CI/CD). NLWeb — Microsoft’s ‘HTML for the agentic web’ — lets websites add a conversational interface in a few lines of code; every endpoint also functions as an MCP server. Early adopters: TripAdvisor, O’Reilly. Microsoft and GitHub joined the MCP Steering Committee. via Microsoft
- Grok 3 and Grok 3 mini land on Azure AI Foundry. Despite Musk’s lawsuit against Microsoft and OpenAI, Nadella aired a pre-recorded Musk conversation announcing Grok 3 as Microsoft-hosted and Microsoft-billed (two-week free preview). Foundry now hosts >1,900 models including Anthropic Claude, Meta Llama, Mistral, Cohere, DeepSeek and BFL. via Fortune
- Anthropic ‘Code with Claude’ teased for next morning. First-ever developer conference opens 9:30am PT May 22 in SF. Builders read this as the launch window for Claude 4 — confirmed by Opus 4 + Sonnet 4 dropping next day.
Who shipped
OpenAI + io reshaped the hardware story. Mistral reset open-source coding SOTA. LMArena turned commercial. Google and Microsoft ran second days of mega-conferences. Anthropic teased the launch.
Open-source pulse
Devstral + Microsoft’s open-sourced Copilot Chat in VS Code are the day’s open contributions. NLWeb as an open protocol is potentially the most consequential — it tries to make every site agent-readable by default.
Money, infra & hardware
$6.5B for io + $100M for LMArena make this a defining funding day. The io deal signals OpenAI is serious about owning the hardware layer; the LMArena round signals that benchmark infrastructure is itself a venture-scale business.
Quiet corners
Google Jules entered public beta inside GitHub (asynchronous Gemini 2.5 Pro coding agent). Stitch — Google Labs UI design tool from the Galileo AI acquisition — also went live.
By the numbers
- $6.5B / ~$5B / 23% — io deal / new equity / pre-existing OpenAI stake
- ~55 — io engineers joining OpenAI
- 46.8% — Devstral SWE-Bench Verified (open-source SOTA)
- $100M / $600M — LMArena seed / valuation
- ~1,479 tok/s — Gemini Diffusion
- 89.6% / 23.3% / 40.4% — Gemini Diffusion HumanEval / AIME 2025 / GPQA Diamond
- >1,900 — models now in Azure AI Foundry
- Most-mentioned company: OpenAI
Compiled by AI Feed’s editor from verified web sources for 21 May 2025.