Charts

Figures on real data, drawn by hand with the site's own chart machinery. Every number is sourced in the data file that renders it; every plate is addressable and embeddable from the ledger.

Everyone quotes the latency table †. Almost nobody draws it, because on a linear axis you can't. Nine decades separate an L1 hit from a transatlantic packet; the table hides that, the scale is the story.

fig. 1.3.1 · the famous latencies, drawn to scale for once.

source: Dean/Norvig constants, 2020s values · latency.data.ts · colin-scott.github.io†

The price of the best model never falls: it launched at $60 a million tokens in 2021 and has launched at $60 or more ever since. What collapses is the price of any fixed capability once the frontier moves past it: roughly ten times cheaper every year †, and up to 40× a year † for the hardest benchmarks.

fig. 1.3.2 · what a million tokens costs, by capability class.

source: published launch prices · tokenPrice.data.ts · epoch.ai† · a16z.com†

Fifteen years of HTTP Archive crawls †, one line: the median mobile page is eighteen times heavier than in 2011, and its JavaScript alone now outweighs the entire median page of 2014.

fig. 1.3.3 · the median mobile page, gaining weight since 2011.

source: HTTP Archive report medians (mobile p50) · webWeight.data.ts · httparchive.org†

Andrew Fribush, software and computer engineer in Austin. I build consequential systems where software meets institutions: finance, healthcare, commerce, government data, and the mechanisms that govern AI agents. Taking new engagements: hello@andrewfribush.com.