Dianzinao

About

Dianzinao Lab studies how people and AI come to understand one another. Two questions guide us: how can people express and confirm what they mean, and how can they gain the knowledge they need to judge an AI’s answer?

Intent alignment

An AI can give a sophisticated answer even when someone is still looking for the right question. If you want an object to “go around here,” tracing a path with your hand may say more than describing each turn. But the gesture leaves something open: does the object need to avoid an obstacle, or pass through a particular point?

We study the exchange that follows. A person might speak, draw or gesture. The AI should be able to sketch the path it understood and point out where it is unsure. The person can then confirm it or make a correction. We call this two-way process intent alignment.

Knowledge synchronization

Sometimes a missing concept gets in the way. You might want to know how fast a curve is changing at one point. That is already a precise question, even if you have never studied derivatives. An answer consisting of a formula alone gives you little basis for deciding whether it addresses what you meant.

By knowledge synchronization we mean helping someone reach the concepts they need, starting from what they already understand, so they can ask further questions and judge an answer for themselves. We sometimes imagine learning calculus by eating a single sweet. It is a thought experiment: if understanding took less time, what questions would people finally be able to ask?

Current work on multimodal interaction

Our work today focuses on interactions people can try. With YuHeng you can talk an idea through. YaoGuang gives you a shared drawing board for a conversation. These prototypes help us ask when an agent should ask a follow-up question, when a drawing resolves an ambiguity, and how gestures and body movements might become part of that exchange.

Longer-term ideas and the name Dianzinao

The following three ideas are long-term possibilities. The lab has not started projects or experiments in any of them.

Brain–computer interfaces for intent. Could a person convey an intention before finding the words for it? How could the system show its interpretation to that person for confirmation?

Direct access to knowledge. Could knowledge one day be conveyed directly through a brain–computer interface? Even if information could be transmitted, how would we know whether the recipient understood it?

The cyberbrain. Dianzinao means “electronic brain,” a name inspired by Ghost in the Shell. It holds our most distant ambition: if human cognition could one day continue in an electronic medium, how might that change the way we explore deep space?

Contact the lab

We would like to hear about problems you found hard to explain to an AI but easy to show with a drawing or a gesture. You can reach us by email.

zhangming@dianzinao.com