Overview
Introduction
Super Intelligence is an open project on one question: how can machine intelligence be advanced to superhuman level safely, and how can what it produces be shared widely. This documentation sets out the research and engineering agenda behind that question.
What this documentation is. A technical agenda and a reading of the field as of October 2026. It describes problems, methods and proposed practice. It does not describe a finished system. See Project status for what exists today.
Who it is for
- Researchers who want a compact map of the open problems between today's frontier models and superhuman systems, with primary sources.
- Engineers who build training, inference, agent and evaluation systems and want concrete practices.
- Reviewers and partners who need to judge whether the project's claims are careful and its methods sound.
How it is organised
| Section | Question it answers |
|---|---|
| Research | What has to be solved scientifically, and which directions are most promising? |
| Safety | How is a more capable system kept aligned, inspectable and under control? |
| Engineering | How should the software and infrastructure be built? |
| Network | How can the output of advanced AI be shared, and what is the token for? |
| Reference | Roadmap, glossary and the full source list. |
Conventions
Every page follows the same order: the state of the field, the open problems, then recommended practice. Three rules apply throughout.
- Claims carry sources. All sources are collected on one page: References.
- Proposals are labelled. Text under a heading such as "Recommended practice" or "Proposed design" is the project's position, not an established result.
- Uncertainty is stated. Where the evidence is thin, disputed or self-reported by a company, the text says so.
Where to start
- For the argument in one page, read the Thesis.
- For the scientific agenda, start at The capability frontier.
- For system design, start at Reference architecture.