New AI on May 20, 2026. The most interesting dated record was not a model. It was a phone learning to pick which chat window to talk to.

Why a dated query and a trending feed disagree

A trending score is built to surface popular things, not new ones. The same model can hold the Hugging Face top slot for a week, and a paper that arrived three days ago can outrank a release that landed this morning. So when the question is literally "what dropped on this exact date," a homepage feed is the wrong instrument. The right instruments are project changelogs and per-day commit logs. They are not as social, but they carry timestamps you can trust.

That is why the rest of this page narrows in on one project. A list of ten trending repositories would tell you what is loud this week. One project's May 20 tells you what shipping actually looked like that day, including the small decisions a trending feed will never index.

One open-source repository's actual May 20

Fazm is an open-source macOS app that wraps Claude Code (through the Agent Client Protocol) and Codex in a native UI, with persistent sessions, one-click forking, and no auto-compacting of context. The repository lives at github.com/mediar-ai/fazm, and on 2026-05-20 it ran through three full release cuts in a single working day. The density of that day is the texture a trending feed cannot show:

The three releases are not a single feature broken in pieces. Each one ships against a different failure mode: v2.9.26 fixed workspace state for pop-out windows after restart, v2.9.30 rebuilt the phone web chat into a full session client, and v2.9.31 reworked the personal-account flow and fixed Python MCPs on universal-DMG installs. Skipping a day was not an option; three users were already running each of those edges in production.

The headline change: a phone learning to address pop-outs by name

Fazm lets a single Mac run many concurrent conversations as detached pop-out windows. Each pop-out has its own workspace, its own agent backend (Claude Code or Codex), and its own model. The problem before May 20 was that the companion phone web chat at chat.fazm.ai had no idea any of those existed. Every phone message routed into the floating bar by default, which meant that on a busy day you could only talk to one of your conversations from the phone.

Three commits landed end to end at 10:27 and 11:13 Pacific (verifiable in git log --since="2026-05-20"):

• 0798558f taught the desktop relay to accept pop-out routing and control commands ( Desktop/Sources/Chat/WebRelay.swift, 67 insertions).
• c659db76 wired the matching handlers on the web side so the phone client could request desktop state and the list of open pop-outs.
• c5ed8da1 added the user-facing piece: a 204-line SessionSidebar.tsx that lists every pop-out by workspace path, plus a new chat button, a pop-out button, a stream-target switcher, and model and workspace selectors.

None of those commits land on a trending list. None of them ship a model. Together they make the phone a peer of every pop-out on the Mac, instead of a thin shortcut into one of them.

What the wire looks like now

The phone client and the Mac broker speak over a single WebSocket. The change introduced one field, sessionKey, that pushes the routing decision out of the relay and into the phone's UI. The desktop side now dispatches each message by that key.

// chat.fazm.ai → desktop relay, after May 20, 2026
{
  "type": "chat",
  "target": "detached-3a1b9c2e",     // the pop-out you picked on your phone
  "sessionKey": "detached-3a1b9c2e", // routes to this session, not "main"
  "text": "summarize the diff we were just looking at",
  "model": "claude-opus-4-7",
  "workspace": "/Users/m13v/fazm"
}

// before May 20, 2026 every phone message was implicit "main"
{ "type": "chat", "text": "..." }    // floating bar always

The sessionKey takes one of two shapes: the string literal "main" for the floating bar, or "detached-<uuid>" for any individual pop-out window. The phone subscribes to a popouts_list push so it always knows what is open. If you close a pop-out from the Mac, the next push prunes it, and the phone falls back to main automatically.

The other two releases worth opening

v2.9.30 was the headline, but v2.9.26 and v2.9.31 carry their own dated decisions. v2.9.26 fixed pop-out chat windows collapsing to the same workspace after an app restart, so the assumption that each window keeps its own working directory became real instead of aspirational. It also restored the copy button on the active question bubble, which had quietly disappeared a few versions earlier.

v2.9.31 reworked the personal-account chooser so a user who is already logged into Claude or ChatGPT in one window can hand that session to a new pop-out, instead of being pushed through a fresh OAuth flow every time. It also fixed Python-based MCP servers (Google Workspace, browser-harness, ai-browser-profile) failing to load on universal-DMG installs by resolving both the thinned .venv/ directory and the per-architecture .venv-arm64 and .venv-x86_64 siblings. Universal binaries that bundle a Python venv per architecture are an obscure corner of Mac packaging that basically no other open-source AI agent has had to touch yet.

None of these are model-layer news. All three are the kind of work that lets a wrapper around Claude Code earn the right to call itself an app and not a UI demo.

How to read any open-source AI project's real May 20

The method generalizes to any repository that catches your eye on Hugging Face or GitHub. Skip the README, which is marketing. Open three dated artifacts in order. The changelog at the root of the repository, which is authored by the maintainer and tied to version numbers. The commit log for the day you care about, which shows what landed and at what minute. Then one diff, where the inline comments tell you why.

If the changelog is empty, the commits are one launch push from three months ago, and the diffs carry no comments, you are looking at a project running on momentum rather than maintenance. Fazm's May 20 passes all three checks, which is the only reason a dated question is worth answering with one project instead of a list of ten.