Research
Faults in the instrument
Most of what looked like a fault in our scorer was a fault in the thing measuring it. One automatic check was wrong six times in seven, and we had built it so that nobody could see that.
Research
Where AI makes someone more capable, and where it does not
Our research is applied. It comes out of building real systems, and the question it answers is always the same one: is the person using this better off, and can we show it rather than assert it.
How the research runs
Applied, and circular. The build produces the question.
- A loop, returning to where it started.
- Build something real, with a constraint it must hold. Find the conditions under which it stops working. Publish the result and its failure conditions together. Change the build, which produces the next constraint.
- The research is applied: the build produces the question rather than following from it.
01 — Build
The work starts with something being built, not with a research question. The question arrives when the build meets a constraint it cannot satisfy, which is why our research is applied rather than frontier.
02 — Find the limit
Where it stops working is the finding. Not where it succeeds — every method succeeds somewhere, and the useful part of the map is the edge.
03 — Publish it
The limit is written down and published with the result rather than after it. If a note has no account of what it does not cover, it is not finished.
04 — Change the build
What we learned changes the thing that taught it to us, which produces the next constraint. This is why the research and the products are not separable, and why we do not publish on an academic calendar.
04 returns to 01
Published notes
Every result carries the conditions under which it fails.
Our commitmentsFollow the streamRSS