AI-NATIVE SERIES · WEARABLE AI · THE WEARABLE LADDER, EPISODE 3 OF 3

AI Glasses Replace the Desk Only When Five Things Come True — Here Are the Dates

By Paul Jialiang Wu · agentic-portfolio-lovat.vercel.app · 2026-09-30 · 14 min read

The one question this episode answersWhat has to become true, and by when, before glasses can replace a desk?

This piece does not answer why earlier head-worn computers failed — that is episode 1, the death table — or what today's glasses can do on a real day, which is episode 2, the documented surface and the pre-registered road-trip audit.

What this AI glasses prediction ledger covers, and who it's for

This is the last episode of The Wearable Ladder. Episode 1 coded every head-worn computer that shipped and failed, and found that 19 of 23 died of money. Episode 2 judged 18 of my laptop tasks against what the glasses are documented to do, and pre-registered the road-trip audit that will check it. This episode looks forward, and does it in the only form that can be graded: a list of specific claims about AI glasses, each with a date, a falsifier, and a number.

It is for anyone who has to decide when — when to build a product for glasses, when to fund one, when to buy a team a pair and expect real work from it. It was written two days after Meta Connect 2026, and every baseline number carries its source and a tag: FETCHED means the number was read on that page; UNVERIFIED means it was seen only in a snippet, a secondary report or a rumour.

The series has one picture, and this episode uses the part of it that looks forward.

The picture: a fight camp with weigh-ins

A fighter who says "I'll be in shape eventually" is not in a fight camp. A camp is a calendar of weigh-ins: on a given morning you step on a scale, the number is the number, and nobody argues with the scale. You make weight or you don't. Before a fighter can turn professional, five things have to be true at once — and a fighter who is great at four of them still doesn't get the fight. So the picture carries two maps at once: each weigh-in is one prediction, and each of the fighter's five skills is one condition.

The thesis: the laptop becomes the server

Here is the claim the whole ledger hangs on. Glasses will not replace a desk by becoming a laptop on your face. The face is the wrong place for a laptop — episode 1's table is a graveyard of devices that tried to put the expensive part up there. Glasses replace a desk when the laptop stops being where you sit and becomes where the work runs, and the glasses become the corner's voice in your ear: you say what you want, an agent with the laptop's full power does it, and you hear what happened.

Connect 2026 put both halves of that on stage in the same week, without joining them. Meta's new agent, Muse, has a Mac app with computer use, and Muse is "coming soon" to the glasses, woken by saying its name (Meta). Separately, the developer toolkit for the glasses reached version 1.0 and added voice invocation — but it only launches an app; it cannot pass the request along (changelog). The corner exists and the fighter exists. Nobody has yet let the corner talk to the fighter mid-round. That gap is prediction P5.

Five conditions, and where each stands today

1. Power — all day on the face. Meta Ray-Ban Display, the display pair, weighs 69 g (Meta) and runs up to 6 hours of mixed use per charge (Meta help). A workday plus a commute is 10 hours or more — my assumption, not a measurement. Episode 1 found that battery and heat were the primary cause of 0 deaths, so this is not a killer — it is a ceiling on how much of the day the glasses can hold. Prediction P1 asks for 12 hours at 70 g, with a camera and colour display, by the end of 2028.

2. Input — writing, not just commanding. Voice is a jab: fast, and good for setting up the round, but you cannot write a design doc by talking to a room. Meta's wristband paper in Nature measured handwriting with the muscle-reading band at a median of 20.9 words per minute in closed-loop use (Nature, 2025) — a useful speed for a reply, and below the 36 words per minute the same paper cites for a phone keyboard. Prediction P2 asks for a median of 40 words per minute, published or shipped, by the end of 2028.

3. Display — a screen worth working on. Today's display pair shows a small, sharp picture in one eye — a monocular display at 42 pixels per degree (Meta). Enough for a message and a map; not enough for a spreadsheet. Prediction P4 asks for see-through colour in both eyes, with a camera and no cable, from a major maker at $999 or less by the end of 2028. Meta's new VR Glasses do not count: they are $1,299.99, they cover your eyes, and they run from a battery pack.

4. Openness — anyone can ship an app. This is the condition people forget, and the one with the most concrete slip on record. At Connect 2025, Meta targeted open publishing for 2026 (Meta). At Connect 2026 the toolkit reached 1.0, but every new capability is marked experimental, and apps using those capabilities "cannot be published yet" (changelog); two curated places to find apps are "coming soon" with no date (Meta developer blog); and the iOS guide still says publishing to the App Store "is not currently supported" (Meta). A fighter who says he'll make weight "this year", then quietly deletes the year, is telling you something. Prediction P3 gives it until 30 June 2027.

5. The laptop becomes the server. The corner has to be able to act. Prediction P5 resolves true only if a request spoken or written on the glasses makes the maker's own agent operate your Mac or PC to finish a multi-step task, generally available, by 30 September 2027.

The ledger

Read the odds as a set. The two I am most confident in are the least interesting: that open publishing arrives within six months of Meta's own deadline (P3, 65%) and that nobody bans wearing camera glasses in public outright (P8, 92%). The ones that decide whether glasses replace a desk — all-day power, fast writing, a real display — are each below even odds by the end of 2028. Three of my five conditions are under 50%: by my own odds, this fighter is still an amateur. Multiply them and the honest answer to "when will glasses replace my desk" is: not in this ledger's window, unless the corner does most of the fighting. Which is the thesis — with one honest caveat: even if the laptop does the work, you still have to tell it what you want, which is why input (P2) stays on the list when the display and the battery matter less.

Five whys: why can't glasses replace a desk in 2026?

1. Why can't you do a desk's work through glasses today? Because the work cannot get in or out fast enough: no third-party app can yet be published to the public, and the wristband's closed-loop handwriting runs at a median of 20.9 words per minute. Evidence: Meta's iOS integration guide; Nature, 2025.
2. Why is the pipe so narrow? Because desk work is mostly producing and reading dense text, and the glasses offer a small display in one eye and a wristband that writes below the 36 words per minute the same paper cites for a phone keyboard. Evidence: Meta Ray-Ban Display (monocular, 42 pixels per degree).
3. Why not just put a bigger screen and a keyboard on them? Because the face limits weight and power — 69 g and up to 6 hours today — and episode 1 shows what happens to devices that put the expensive part on the face: 19 of 23 died of money. Evidence: content/series/wearable-ladder/deaths.yml.
4. So where can the work go instead? To a computer that isn't on your face. Meta already ships an agent that operates a Mac, and has announced it for the glasses. Evidence: Connect 2026.
5. Why hasn't that happened? Because the glasses can only send the laptop an intent if an agent is allowed to receive it, and the platform does not yet let one: voice invocation launches an app without the request, and the connectors to Meta AI are early access. The root: the desk disappears when the glasses carry intents to an agent with the laptop's power — so the bottleneck is the developer surface and the agent's reliability, not the lens. Evidence: toolkit 1.0 changelog; docs/plans/2026-09-25-glasses-connect-2026-delta.md, rows 6–7.

Operate — the ledger runs

The ledger is a file, content/series/wearable-ladder/predictions.yml, and the table above is generated from it. Scoring a prediction is one edit — add an outcome with its evidence:

  - id: P3-open-publishing-meta
    ...
    outcome: {result: false, evidence: https://developers.meta.com/wearables/faq/, scored: 2027-07-01}

Then node scripts/check-wearable-ladder.mjs --metrics reports how many are resolved and, as soon as one is, a Brier score — the mean of (my probability − what happened)², where 0 is perfect foresight and 0.25 is what you get by saying 50% about everything. The claims, dates and probabilities are never edited after publication. You can add an outcome; you cannot move the goalposts.

Feedback — how it goes red, and what it cannot see

Thirty days after a prediction's resolution date, if it has no outcome, the site's build fails: npm test runs the series gate, and the gate names the unscored prediction. That is the whole trick. The repository refuses to deploy anything until I have looked at the scale — and the scale doesn't care that I wrote the odds. It cannot see three things:

• Which predictions I chose. Eight is a selection. A ledger can be perfectly calibrated on easy claims and silent on the one that mattered. The five conditions were fixed first, and each prediction names the condition it tests, so a gap is visible — but the choice was mine.
• Whether my odds are good before 2027. Nothing resolves until mid-2027. Until then the Brier score is empty, and I have said so rather than printing a zero.
• Definition drift. Analyst categories change. P6 depends on IDC's "optical see-through" line; if IDC redefines it, the prediction's text says to use the closest line and note the change.

Feedforward — what I will do when each weigh-in comes in

The point of committing now is to decide the response before the result can bias it:

• If P3 resolves false (no open publishing by mid-2027): stop planning a glasses app for the Meta store. Build for Meta's web apps and assistant connectors, or for another maker's glasses — distribution is the purse, and episode 1 says the purse decides.
• If P5 resolves true (the maker's agent drives your laptop from the glasses): run the episode 2 audit with the same pre-registered task list, again, and compare. That is the test of the thesis.
• If P2 resolves false (no 40-words-per-minute input by 2029): assume writing stays at the desk and design glasses work around approving, routing and listening — the jab, not the knockout.
• If P8 resolves false (a general public ban somewhere): re-open episode 1's social row. It was one death in 23; a ban would make social backlash a live cause again.

Where this episode sits in the series

Same story, five exits

Build it in 30 minutes

• Input: a decision your team is making that depends on the future ("we'll need X by next year").
• Output: three to eight rows, each a claim with a threshold, a date, a public falsifier, and a probability.
• Invariant: claims, dates and odds are frozen when written; only outcomes are added.
• Stop rule: stop when every assumption your decision rests on has a row.
• Minimal schema: id, claim, resolution_date, falsifier, current_evidence[{value, source, tag}], my_probability, outcome?.
• CI threshold: the build fails when a row is 30 days past its date with no outcome.
• Deliberately missing: a second forecaster. Two people's odds on the same rows, scored side by side, is the upgrade.

5S self-score, with evidence

References

1. The ledger and every baseline source: content/series/wearable-ladder/predictions.yml.
2. Meta Connect 2026 announcements: Meta.
3. Toolkit 1.0 changelog: facebook/meta-wearables-dat-ios.
4. Toolkit publishing status: Meta developer blog, Connect recap; iOS integration guide.
5. The original 2026 publishing target: Meta developer blog.
6. The neuromotor wristband paper: Nature, 2025 (PMC full text).
7. Meta Ray-Ban Display specifications: Meta.
8. Market baselines: IDC; UploadVR on EssilorLuxottica.

Related

• Episode 1 — the death table: 23 head-worn computers, and why 19 of them died of money.
• Episode 2 — my 18 laptop tasks, judged by the documentation: 11 should work hands-free; the road-trip audit will check.
• The Assumption Ledger — the same discipline applied to safety claims.