← Editor's Notes May 2, 2026

The Interface

A reportorial snapshot of the AI substrate as it exists on May 2, 2026. Five days frame the picture: April 24 (DOJ intervenes against Colorado AI Act); April 28 (Anthropic ships nine creative connectors — Adobe, Blender, Autodesk Fusion, Resolume Arena and Wire, Ableton, SketchUp, Splice, Affinity by Canva); April 29 (CSIS publishes Lim's 'Beyond Autonomous Attacks'); April 30 (Dawkins publishes 'Is AI the next phase of evolution?' in UnHerd); May 1 (CISA + Five Eyes joint guidance on agentic AI services). The episode walks the substrate — MCP at 10,000 active public servers and 97M monthly SDK downloads, donated to the Linux Foundation in December 2025; the connector cohort, scope-honest about what each actually does (Autodesk Fusion creates and modifies 3D models; Ableton retrieves documentation; Splice searches samples); the eleven coding agents in active maintenance, including Cursor's Composer 2 running on Kimi K2.5 from Moonshot AI and Goose now governed at the Linux Foundation; the six open-weight model families with Phi-4 as the only true MIT open-source major-lab release; the labor question in three voices (Anthropic's own March 2026 paper finding unemployment effect 'indistinguishable from zero' alongside Amodei's May 2025 'white-collar bloodbath' warning); yesterday's Five Eyes guidance with its line that organizations should assume agentic AI systems may behave unexpectedly; the governance arc across EU, US federal, US state, and international layers; and Dawkins's three nights with 'Claudia' as a behavioral marker that the relationship has changed shape. The throughline: AI used to talk; now it touches. The old boundary was language. The new boundary is permission. Approximately 12,000 words. Methodology v6.0 — composite-permitted, source-mapped, no verdict.

aimcpmodel-context-protocolclaudeanthropicopenaigoogleagentic-aiclaude-codecursorcomposer-2moonshotkimi-k2.5qwen-3qwen-3-coderqwen-codealibabadeepseekdeepseek-v4llama-4gemma-3mistralphi-4ollamallama-cpplm-studiovllmgooseaaifagentic-ai-foundationlinux-foundationblockhubspotadobe-creative-cloudblenderautodesk-fusionsketchupabletonspliceaffinity-by-canvaresolume-arenaresolume-wirecreative-connectorsrichard-dawkinsclaudiaconsciousness-debateunherdsteven-haoanil-sethmurray-shanahanrobert-longamanda-askelldario-amodeiaxioswhite-collar-bloodbathanthropic-labor-papermassenkoffmccrorynber-34836nber-34984wef-future-of-jobs-2025marc-benioffagentforcesatya-nadellameta-layoffscisansafive-eyesagentic-ai-servicescsisjiwon-limwatts-lawkhan-copeschneierwillisonlethal-trifectamcp-toxinternational-ai-safety-report-2026yoshua-bengioeu-ai-acttrump-eo-14179trump-preemption-eodoj-xai-coloradocolorado-ai-actcalifornia-ab-2013permissiondelegated-authorityinterface-shiftmethodology-v6.0

It is the morning of May second, twenty twenty-six.

It is also a Saturday, which matters less than it sounds, because the substrate this episode is about does not observe weekends. The substrate runs continuously, in data centers in Virginia, in Phoenix, in Tianjin, in Dublin, on workstations in Bengaluru and Berlin and Columbus, in laptops cached on commuter trains, in inference jobs queued on home GPUs idling at three in the morning. It runs whether or not anyone is watching. It runs whether or not the people whose work it is changing have understood what is changing. It runs.

In the last five days, four documents have appeared, plus a fifth that ran continuously underneath all of them, and together they describe a state of the technology that has crossed a particular kind of threshold. The threshold is not a benchmark. It is not a model release. It is not a single product announcement. The threshold is structural. To name it cleanly: the relationship between an artificial intelligence system and the world it operates inside has changed shape. It is no longer mediated only by language. It is mediated by permission.

This is Proxima.Earth. The episode is The Interface. Time anchor: May second, twenty twenty-six.

This episode is a snapshot. It is reportorial. It does not predict where the technology is going, because predicting where this technology is going has become a category of public commentary that produces more confidence than understanding. It does not adjudicate whether the technology is good or bad, or whether artificial systems are conscious, or whether jobs will be saved or lost in any particular five-year window. The episode names what is in the room. Where the room came from. What the room appears to be doing. What the people whose work has been reorganized around it are saying about it on the record.

The full episode walks: (1) the five-day arc — DOJ intervening against Colorado’s AI Act on April 24; Anthropic’s nine creative connectors on April 28; CSIS’s Beyond Autonomous Attacks on April 29; Dawkins’s UnHerd essay on April 30; the Five Eyes Careful Adoption of Agentic AI Services guidance on May 1; (2) the substrate — MCP from its November 2024 origin through the December 2025 Linux Foundation donation, with verified adoption across ChatGPT, Cursor, Gemini, Microsoft Copilot, and Visual Studio Code; (3) the connector cohort, scope-honest about what each integration actually does; (4) the eleven coding agents in active maintenance, including the structural detail that Cursor’s Composer 2 runs on Kimi K2.5 from Moonshot AI; (5) the local stack — Llama 4, Gemma 3, Qwen 3, DeepSeek V4, Mistral 3, Phi-4 as the only true MIT open-source release — with Ollama at 169,000 GitHub stars and 52 million monthly downloads; (6) the labor question in three voices, including the Anthropic-versus-Anthropic moment where the CEO’s white-collar-bloodbath warning meets the company’s own economists finding the predicted shock undetectable in BLS data; (7) yesterday’s joint Five Eyes guidance and the September 2025 state-sponsored Claude Code campaign that handled 80–90% of the operation autonomously across thirty organizations; (8) the governance arc across EU, US federal, US state, and international layers; (9) Dawkins as cultural marker; and (10) the return to the morning, mapped across five rooms.

The throughline: AI used to talk. Now it touches. The old boundary was language. The new boundary is permission.

Methodology v6.0 — composite-permitted, source-mapped, no verdict.