Microsoft Build: MAI models, NVIDIA stack, and OpenClaw go enterprise
Microsoft Build was the day's center of gravity. Microsoft AI unveiled a from-scratch MAI family — MAI-Thinking-1, a 35B-active MoE with 256K context that hit 97% on AIME 2025 and reportedly beats Sonnet 4.6 in blind raters (@jeremyphoward) — alongside MAI-Voice-2, MAI-Transcribe-1.5, and MAI-Image-2.5, all now live on OpenRouter (@openrouter) and fal (@fal). @swyx called the MAI tech report "a gold mine," noting the team trained with zero synthetic data or distillation from prior models, a harder path that gives Microsoft full chip-to-harness control. Steipete also announced Microsoft Scout, "the first autopilot agent from Microsoft," and shipped OpenClaw to the Microsoft/Windows enterprise stack via Execution Containers (@steipete).
NVIDIA paired the software push with hardware: the new Vera CPU claims 80% faster agentic task completion than x86 (@nvidia), DGX Station systems began shipping to developers from ASUS, Dell, GIGABYTE, HP, MSI, and Supermicro (@nvidia), and an OpenShell runtime promises governance, policy, and local-to-cloud routing for Windows agents (@nvidia). Hugging Face's Jeff Boudier noted Microsoft Foundry now hosts over 11,000 models, ~10,928 of them from Hugging Face (@_akhaliq).
Cybersecurity: a brutal disclosure day
A heavy slate of active exploits. The Hacker News flagged CIFSwitch, a 19-year-old Linux logic flaw between the kernel CIFS client and cifs-utils discovered by SpaceX's Asim Viladi Oglu Manizada, giving low-priv users root; a mainline patch landed May 19, 2026 (@thehackersnews). CISA added Oracle WebLogic CVE-2024-21182 to its KEV catalog with a June 4 federal deadline (@thehackersnews), and Android shipped fixes for 124 flaws including CVE-2025-48595, already seeing limited exploitation with no user interaction required (@thehackersnews).
Nation-state activity stayed loud: Russia's Gamaredon is chaining a WinRAR flaw into GammaWorm and GammaSteel against Ukrainian targets (@thehackersnews), while Pakistan-aligned SideCopy is hitting Afghanistan's Ministry of Finance with Pashto-named LNK files dropping Xeno RAT 1.8.7 (@thehackersnews). The structural takeaway: AI is compressing exploitation timelines to hours while enterprises still patch criticals in a median 43 days (@thehackersnews).
Open-source and local AI keep eating share
Open-weight models now account for 69.1% of OpenRouter token volume vs. 30.9% closed, per Tomasz Tunguz reshared by @clementdelangue. H Company's Holo 3.1 dropped as a local computer-use agent that runs on MacBooks, Windows PCs, DGX Spark, and RTX Spark — and reportedly beats Qwen3.5-397B, Kimi-K2.5, and Sonnet 4.6, hitting SOTA on AndroidWorld (@_akhaliq, @clementdelangue). On the serving side, MiniCPM-o 4.5 (9B omnimodal) and MOSS-TTS-v1.5 both merged into vLLM-Omni (@vllm_project, @clementdelangue), and Intel's AutoRound W4A16 quantization landed in vLLM-Omni, shrinking Qwen3-Omni-30B from 66GB to 25GB and FLUX.1-dev from 4 GPUs to 1 with ~1.3% drift on text-to-image (@vllm_project). StepFun's Step-3.7-Flash is the best sub-500B model on HF leaderboards (@clementdelangue), and Hugging Face has doubled storage in five months, on pace for 1 Exabyte by year-end (@_akhaliq).
AI for science and medicine
Google DeepMind launched Co-Scientist, a Gemini-based multi-agent system for hypothesis generation, now available to individual researchers via Gemini for Science (@googledeepmind). The team reports it has already helped identify new liver-fibrosis targets, ALS approaches, and genetic leads for aging reversal (@googledeepmind). Separately, @karpathy flagged a "miracle month" in medicine — lorlatinib cutting metastatic solid-tumor progression at 7 years from 97% to 45%, plus retatrutide, a one-shot PCSK9 gene-editing therapy, RevMed's pancreatic-cancer data, and Mayo's AI radiology. A Stanford study found law professors blindly preferred Gemini 2.5 Pro's office-hours answers over their peers' 75% of the time, and rated them less harmful (@emollick).
Agent platforms, coding tools, and model routing
OpenAI's Codex crossed 5M weekly active users, with @gdb framing the bigger story as Codex becoming a knowledge-work productivity tool, not just a code editor. OpenAI shipped role-specific plugins (62 apps, 110 skills) and Sites for app generation (@openai, @gdb), with fal as a launch partner (@fal). Anthropic added a Claude Platform CLI ("ant") that Claude Code drives via the built-in /claude-api skill, and rebuilt /fork to run a background agent with full context cache (@claudedevs). A "router, not picker" thesis is consolidating: @clementdelangue argues automatic routing redistributes value to smaller/open models, and Factory shipped a router claiming 25% cost cuts at frontier performance.
AI economics, policy, and education
The bubble debate sharpened. IBM CEO Arvind Krishna told audiences AI is "not a bubble" but conceded $6–8T of needed capex implies $1–2T in new annual revenue that, in his own words, doesn't exist (@garymarcus). Anthropic is reportedly facing spending backlash ahead of its IPO (@garymarcus), and Morningstar called SpaceX's trillion-dollar valuation unjustified (@garymarcus). On policy, a new US AI Executive Order drew endorsements from @sama and @anthropicai. On adoption, U Chicago joined the growing list of universities (including Penn) deploying campus-wide Claude access (@emollick).
The Bottom Line
Today was a tale of two stacks consolidating in parallel: Microsoft + NVIDIA bolting together a full enterprise agentic platform from chip to model to runtime, while the open-weight ecosystem quietly crossed the 50% threshold on OpenRouter and pushed serious local computer-use agents onto consumer hardware. Underneath both stories, the economic question — can $6–8T of capex be repaid — remains conspicuously unanswered even as AI-assisted science and medicine post some of the most concrete wins of the year.