About
Our mission is to create a machine that teaches itself.
In essence, a new type of autodidact.
autodidact /ˌɔː.təʊˈdaɪ.dækt/ · noun
From Greek autodidaktos, "self-taught."
- A person who has taught themselves, rather than having lessons.
- One who has learned how to learn, and so can go on learning anything.
- A machine that teaches itself. (our new definition)
MAD stands for Make All Discoveries. #
We think every problem we haven't solved is one we haven't yet thrown enough intelligence at, and a machine that teaches itself to solve problems will generate enough intelligence to solve any problem.
Goals #
- Create a universal reasoning engine: one machine for any problem.
- Keep knowledge separate from reasoning, so both can be inspected.
- Match frontier models with 100 times less data and compute.
- Learn and run on the device you already own.
- Reach recursive self-improvement: get better at getting better.
- Make every change it makes to itself visible and reversible.
Non-goals #
- Products, customers or revenue.
- A machine that knows everything, trained on the whole internet.
- Bigger models and more compute.