Travel Tech | Lee C. Zhao, MD

What do you actually do with the Vision Pro?

Two jobs. In the operating room it is a camera. Robotic cases record themselves at the console, so the headset earns its keep in open reconstruction, where the operative field is wherever my head is pointed. It captures the case from the surgeon's eye line, a view no boom camera or circulating photographer has ever gotten me.

The open cases in my video library were shot this way. Three of them are below. One runs 50 minutes uncut, which is what a case looks like before editing makes it look easy.

Recorded on the headset

For the rest, open video.leezhaomd.org and type Vision Pro into the search. The headset footage sorts itself out.

On the road the job flips: the headset becomes the monitor. It extends my laptop screen, which turns any seat with legroom into a workstation with more display than my office has, and it does that in a conference room without taking a second screen out of the bag.

Lee Zhao seated at a conference table wearing an Apple Vision Pro headset, working at a laptop among other attendees in a meeting room — The headset, on. Yes, it looks like that.

The screen extender from a seat at the BAUS FNUU and AGUS meeting, March 2025: the Mac display floating over the room, the programme open. A few windows are blurred on purpose.

The cost is storage. Four open cases produced about 40 GB of footage, and a single continuous take can pass 7 GB. Almost none of it survives. A two-minute clip of the right dissection is worth more in a lecture than the ninety minutes around it, which makes the headset a strange instrument: it films everything so that almost everything can be thrown away.

Are those the Meta glasses?

Yes. Ray-Ban Meta Wayfarers, bought at the Ray-Ban store in Florence in June 2025, with Transitions lenses so one pair covers indoors and out. Camera in the temple, open-ear speakers, microphones, Meta's assistant when asked. Of everything on this page they require the least explanation, which is most of their charm.

Most of their work happens on runs. Audio comes through the open-ear speakers, so nothing sits in my ears, and the button on the temple takes a photo of whatever I am looking at without breaking stride. The phone stays in the pocket.

A wide river at dusk under a clear blue sky. A long bridge sits low on the left horizon and a skyline of tall towers stands across the water, both in silhouette against a pink and orange afterglow. In the foreground a lit lamppost and leafy branches overhang a railed stone walkway. — Shot with the glasses, not about the glasses.

Shot with the glasses, on Fifth Avenue: the Louis Vuitton building dressed as a stack of trunks.

Shot with the glasses at work: ten seconds of an open case from the temple camera.

What are the glasses with the green display?

Even Realities G2. They pass as ordinary frames. The display is a 640 by 350 monochrome green panel only I can see, a window about 27 degrees wide in the middle of the view. The constraint is the point: too small to be a screen, exactly big enough for an answer.

A pair of Even G2 glasses, dark matte frame with the temples open, resting on a laptop keyboard in front of a screen showing turquoise water and boulders. — The G2. The display is invisible from your side.

What it does

Out of the box it does what Even lists: a teleprompter, live transcription and translation of the person across the table, a heads-up display for what the phone would otherwise interrupt me with. Each one is a line or two of green text laid over whatever I am already looking at. There is nothing to hold or look down at, and from the other side of the table there is nothing to see.

First-person view through the Even G2 glasses, the dark rims of the frame faintly visible at the left and right edges. Across a round wooden table in a hotel lobby, a man in a navy blazer is mid-sentence, one hand raised; a coffee cup, a notebook with a pen, and a glass of water sit on the table. Floating in front of his chest, small lime-green text: a rounded Prep Notes pill at the upper left, a microphone icon and 11:20 at the upper right, two lines of transcript reading "...so for the long-gap redo we ended up needing a graft, and the question was which one", then "Concept Buccal mucosa graft: lining from inside the cheek, used to rebuild the urethra". An AI-generated illustration, not a photograph. — Conversate from my side of the lenses. An AI render, not a photograph.

The G2 and the ring in use: Even's own launch video. Open on YouTube.

The ring is the control panel. Even's R1 has a touchpad for a face: tap to select, scroll to page through whatever the display is showing. My hand never goes to the temple of the glasses, so from across the table the whole exchange looks like a man fidgeting with a ring.

The control panel. It is also a ring: thumb on the face, tap and scroll.

The part worth explaining is behind the wake word. "Hey Even" does not go to a company's assistant. It goes to a small bridge server I run on an always-on Mac at home, reachable from anywhere over Tailscale. Change one URL in the phone app and the same glasses answer from Gemini, ChatGPT, or Claude. Gemini holds the everyday slot.

The bridge is capture-first, which is the feature I actually wanted. Open with "note" or "remember" and the words are written into my Obsidian inbox before any model is called, so a dropped connection can cost me an answer but never an idea. Open with "research" and it saves the question, answers with search grounding, and lands the full cited answer as a note waiting at home. Deep tasks go to the Claude slot and deliver later over Telegram. One failure mode remains: if the Mac at home sleeps, the assistant sleeps with it.

What is your AI stack?

The honest answer is unglamorous: the load-bearing part is not a model, it is a folder. Everything durable lives in an Obsidian vault, plain Markdown files in a private git repository. Voice gets things in: Wispr Flow for the five-second dictation, a Plaud recorder for meetings and walking brainstorms. The models are clients of that record, not homes for it. Model memory is a convenience cache; the vault is canonical.

The routing, as of this writing: Claude reads the vault and does the synthesis, drafting, and judgment work. Codex executes: code, browser work, bulk verification. Gemini and NotebookLM handle source intake. Grok watches X. ChatGPT is kept as the outside reviewer, deliberately independent of the work it critiques. Rankings shuffle monthly, so the routing rests on what each system uniquely touches, not on this week's benchmark.

One rule makes the whole thing usable: every citation, number, and clinical claim that comes out of a model is treated as wrong until verified against the source. A model is a fast resident who can draft and organize. It is not the attending of record.

How do you edit your videos?

The question assumes the bottleneck is an editing app. It is not. The bottleneck is the library: 9,795 video files accumulated in Dropbox over the years, many of them full-length operative masters. The current working file runs 93 minutes. One older master runs past six hours. Nobody edits their way out of that by opening a timeline.

So this year the first move was cataloging, and I made the machines do it: AI agents walked the entire tree, built a manifest, deduplicated files by content hash, classified what remained, and staged the teaching material for YouTube. Robotic cases record themselves at the console; open cases come off the Vision Pro. The edit itself is mostly subtraction, deciding which two minutes of a ninety-minute case earn the room. Filming is easy. Deleting is the skill.

Watch the result

The finished videos live on video.leezhaomd.org and the @lee_c_zhao YouTube channel. The open cases recorded on the Vision Pro are listed under the headset. Three robotic cases, recorded at the console: