Inventions
Inventions
I spend my days trying to make machine learning checkable. The idea is simple to say and hard to do: a result should be able to carry its own proof, so that someone who was not in the room can still confirm it later.
The problem I keep running into is not that models are wrong sometimes. It is that almost nobody can tell when they are. Verifiability, auditability and reliability are not features you bolt on afterwards; they have to be part of how the system runs in the first place.
- 1
A system runs
- 2
It produces evidence of what it did
- 3
Anyone can check that evidence later
- Verifiable computation for AI
- Provenance and audit that a third party can run
- Cryptographic evidence that outlives the run
Patents
Behind the scenes, that work has become a growing patent portfolio: close to 76 filings in the pipeline for verifiable AI infrastructure, twenty of them filed so far, and more on the way.