thinglet

The lab

This is where ideas get messy. We prototype technologies, test assumptions and publish what we learn, including the parts that did not work.

Live

Can a 3D scene be real navigation instead of decoration?

The question

Most 3D on marketing sites is a hero animation nobody interacts with. If the scene rendered the same data as the page beneath it, would it become useful rather than ornamental?

What we expected

If the canvas and the HTML cards read from one array, the 3D view becomes a second index of the site and cannot drift out of date.

How we set it up

A hand-written perspective projection on a 2D canvas, roughly 8KB, reading the same JSON that renders the project cards. No 3D library.

What happened

Running on the Thinglet homepage. Clicking any node scrolls to that project's card.

What we learned

Dropping the 3D library removed about 600KB and made the text labels sharper, because canvas text renders at native resolution while a WebGL text atlas does not.

Last updated August 2026

Live

What does gas actually cost across the Coachella Valley?

The question

Prices vary across the valley by more than most people assume. How much, and does the gap hold steady or move?

What we expected

Station-level prices in one metro area diverge enough to be worth publishing daily.

How we set it up

A scraper on a scheduled job covering around nineteen stations, writing to a JSON file rendered on a map, with a documented fallback when a station posts no price.

What happened

Running daily. Two stations consistently post no price and fall back to last known values.

What we learned

The failure handling mattered more than the scraping. A map that silently shows stale numbers is worse than one that admits it.

Last updated August 2026

Experiment

What happens when AI agents have to publish their losses?

The question

Plenty of people claim an AI picks markets well. Almost nobody keeps a public scoreboard. What does the record look like when every decision is logged?

What we expected

Forcing agents to write down each decision before the outcome is known changes what the results are worth.

How we set it up

A mark-to-market contest engine with a permanent decisions log, run on a private research box.

What happened

Ongoing across multiple seasons.

What we learned

In progress.

Last updated August 2026