Perspectives

What we think about AI, and what we are still working out.

Nobody has AI figured out, us included. This is where we work in the open. We would rather leave a question standing than publish an answer we cannot stand behind yet.

We are rooted in the history of this field, and we take its risks as seriously as its promise.

The questions

Six things we are trying to get right.

One card each, in plain words. Where the page behind a card is written you can read it. Where it is not, the card says so and stops there.

HistoryWritten: Where did any of this actually come from? It starts with a mathematician counting the letters of a poem to settle an argument, runs through a century of the field trying almost everything else, and ends up back at language.
LanguageWritten: Why does a machine that only predicts the next word turn out to be useful? Our answer so far is that language is not a tool people use but the system they run on, and a machine trained on the written record inherits that system. What else it inherits is the part we are still working out.
ImpactIn progress: What AI gives you arrives quickly and personally. What it costs turns up later, somewhere else, and often on someone who was never at the keyboard. The question is how to hold both on the same surface without averaging them into a verdict.
Capital & computeIn progress: Where is the money actually going, and what is the build waiting on? We think it has moved out of software and into buildings and power, and that the pace is set by things harder to see than chips. Who carries that cost while it waits to pay off is the open question.
Sovereignty & rulesIn progress: Who decides what you are allowed to do with this? Not any one country, as far as we can tell: the chips, the machines to run them and the rules that govern them sit in different hands, pulling different ways. Where those hands are tells you more than which model is ahead.
How we knowIn progress: Almost nobody who quotes a number about AI counted it. A few institutions and researchers do the counting, everyone else copies, and every copy is a chance for the number to go wrong. This is where we write down our own trip back to whoever counted.

Our position

Where we stand.

Two people run this house, and the rule that governs these pages governs what they build.

Deep respect, in both directions. What AI does for people and what it does to them are both real, and neither cancels the other. We build to reduce the second, and we say so before we say anything else.

what it changes in how we build

This house is two people, and it runs on the systems these pages describe, so the rule that governs the pages governs the products. A figure without its source and the date it was true does not ship, here or in what we build. Where the evidence is thin, the product says the evidence is thin, the same way this page does. When we are wrong, the correction stays on the record.

Kept by two people, not by a research department. Devin Sawyer and Jason Schwartz.