What we study, and how we publish it
The questions Hyuge investigates, how it publishes, and how researchers can work with it.
Research areas
Four questions we're investigating
Learning from experience without pre-training
Can a system start from nothing and become useful by learning from each piece of data as it arrives? We study what it takes for learning to keep going, without a fixed corpus collected in advance and without starting over each time something changes.
Calibrated confidence, and knowing what you don't know
When a system says it is sure, how often is it right? We study how a system can report confidence that matches its real record, and how it can recognise when it lacks the information to answer and say so.
Using private information without exposing it
Can private information shape an answer while staying under its owner's control? We study how to keep what information means separate from who can see it, so sensitive data can help without being handed over.
Meaning that persists over time
Why should what a system has learned be thrown away each time a model is replaced? We study how meaning can be kept and built on over years, so knowledge accumulates instead of being rebuilt.
How we publish
Openness is the default
- We publish our findings. Holding work back is the exception, and it has to be justified.
- Every finding is published with a stated confidence level, so readers know how much weight it can bear.
- We withhold work only when it falls on a short, published list of catastrophic risks. When we do, we publish that we have withheld something and why.
Publications
Published work
Our first publications are in preparation.
Collaboration
Work with us as a researcher
We collaborate with researchers in universities and other labs on the four questions above. Collaborations can be a shared study, a visiting position, or a review of our published work.
Tell us what you work on and where you think it meets ours. A researcher on our team will reply.