Advancing intelligence beyond human expertise.
We pursue intelligence that can extend what people are able to discover.
By superintelligence, we mean systems whose intellectual capabilities substantially exceed those of the best human experts across nearly all major intellectual domains. That ambition gives us demanding questions to investigate. Can a system recognize when its assumptions fail? Can it carry an unfamiliar problem through a sustained investigation, choose an informative experiment, and produce a result that withstands expert scrutiny?
Progress can begin with a better learning method, a physical insight, or a new way to test an explanation. We develop technologies and build companies that bring such advances into practical use. The problem and the evidence determine which methods deserve attention.
An intelligent system should carry knowledge across tasks, connect information from different sources, and recognize when an earlier conclusion needs revision. This raises questions about memory, reasoning, and collaboration among specialized systems. The idea of a neural commons belongs here: a shared foundation of knowledge and representations that different systems can develop and reuse, with their sources and uncertainty available for examination.
Models let us investigate consequences before committing to an action. Learned world models can capture patterns in experience; physics-informed AI can incorporate physical knowledge into learning and prediction. Co-simulation connects models of interacting systems so that mechanics, heat, software, and control can be examined together. The choice of model should follow the question being asked, and its predictions should be tested against relevant observations.
An intelligent system should be able to question the assumptions behind its own answers.
A useful explanation helps establish what would change an outcome and under which assumptions. Causal reasoning examines those relationships through models, observations, and interventions. In explainable AI, this brings attention to the mechanisms behind a prediction and the evidence supporting them. An explanation should also reveal what further evidence could change the conclusion.
X causes Y
X 0.3
Y 0.3
Y causes X
X 0.3
Y 0.3
Setting X to 0.7: where X causes Y, Y becomes 0.7. Where Y causes X, Y stays 0.3, because setting X removes its dependence on Y.
Discovery involves choosing a worthwhile question and finding a way to test an answer. Advanced intelligence could help connect findings across disciplines, formulate hypotheses, develop algorithms, and explore designs that are difficult to search by hand. Useful progress produces something others can examine: a proof, a reproducible computation, an experimental result, or a prediction tested against new observations.
These questions also connect our work to security intelligence, where decisions must hold under disruption, adversarial behavior, and consequential constraints.
Explore Security IntelligenceHow today's AI could advance beyond human intelligence, and what that would mean for discovery and security.
