The AI Museum

Each era is a room, and the walls change as time advances.

Five wall colours run from aged light brown to near-white, so the Foundation era feels lit and the Agentic era feels clinical. This is the whole argument of the piece delivered without a sentence of copy. It also does structural work: a reader always knows roughly where they are in nine years by the colour of the wall behind them, which no progress bar achieves. Every wall was contrast-checked against every text tone and accent that appears on it, worst case 5.3:1.

An absent parameter count is the exhibit, and is never estimated.

Sixteen of thirty-one works published a parameter count. Fifteen did not. The chart draws the fifteen as unmarked ticks on the axis: present, unmeasured. Filling those with a rumour would have produced a smoother curve and destroyed the only finding the exhibit has, which is that after 2023 every disclosed figure belongs to an open-weight model and not one closed frontier system has stated its size. The measure did not get harder to take. It stopped being offered.

The chart separates its series by shape, not by colour.

Open weights are filled circles, closed models are hollow ones. The two hues chosen first looked clearly distinct and measured only 1.07:1 apart in relative luminance, which means they merge in greyscale and for a red-green colour-blind reader. Colour is decoration here and the shape carries the meaning. The chart also ships a full data table and an aria-label stating the finding in words, because a chart nobody can see should still be readable.

Contrast is computed, never judged.

The dark lobby was the worst offender and looked fine: its muted text measured 4.03:1, and two greys beneath it 2.58:1 and 1.89:1. One pass of a script converting hex to relative luminance and applying the WCAG formula found four failures that no amount of looking had caught. amber-700 was the documented accent for months at 3.93:1 and is now banned project-wide. Nothing here is a matter of taste; it is arithmetic, and it is cheap to run.

Visit memory reads localStorage as an external store.

Tracking which exhibits a reader has seen is the textbook case for useState plus useEffect, and that version is wrong: it renders once with the empty value and then corrects itself, which is a visible flash and a hydration mismatch waiting to happen. useSyncExternalStore models it correctly, and the server snapshot is empty so hydration always agrees. The subtlety is that getSnapshot must return a stable reference when nothing changed. Returning a freshly parsed array each call type-checks, passes review, and re-renders forever.

function getSnapshot(): string[] {
  let raw: string | null = null;
  try {
    raw = window.localStorage.getItem(KEY);
  } catch {
    raw = null;
  }
  if (raw !== cachedRaw) {
    cachedRaw = raw;
    try {
      const parsed = raw ? JSON.parse(raw) : [];
      cachedList = Array.isArray(parsed)
        ? parsed.filter((x): x is string => typeof x === "string")
        : EMPTY;
    } catch {
      cachedList = EMPTY;
    }
  }
  return cachedList;
}

Reduced motion needs two mechanisms, because one cannot reach the other.

A CSS prefers-reduced-motion rule cannot stop Framer Motion, which animates via JavaScript transforms and never consults the stylesheet. The View Transitions API does not consult the preference on its own either. So the museum wraps itself in a motion config honouring the user setting, and separately disables the view-transition animations in CSS. Shipping only the CSS half would have looked correct in a code review and done nothing for the reader who asked for it.

A performance win was refused three times, and the refusal is recorded.

content-visibility: auto would genuinely help the thirty-one catalogue rows. It is still not enabled, because those rows reveal via a viewport-triggered animation, and if the browser defers rendering the reveal may never fire and leave a row stuck at zero opacity. That failure is invisible in server HTML and so cannot be caught by the way everything else here is verified. It goes in after somebody watches it in a real browser, and not before.

4.03:1

The contrast ratio of text that looked perfectly readable.

Found by computing rather than by looking, along with two worse cases at 2.58:1 and 1.89:1 in the same pass. Every text and background pair in the museum now passes WCAG AA. Touch targets went from about 23px to 44px in the same audit, which was also invisible until it was measured.

16 of 31

Works that published a parameter count at all.

The other fifteen are drawn as unmarked ticks rather than omitted or estimated. After 2023, all eight disclosed figures belong to open-weight models. The exhibit is the gap, and it only exists because the data model allows a field to be genuinely absent rather than defaulting it to zero.

43 pages, 3 arrays

Everything is generated from data, nothing is authored twice.

Forty-three routes, the sitemap, the structured data, the homepage timeline and the guided tour all read from the same three TypeScript arrays and forty-two typed JSON files. Adding a work is one array entry plus one file.