People frequently ask Google questions like “What career is right for me?” “How can I find a boyfriend?” or “How can I make more money?” While Google can answer self-contained questions like “What is the capital of Turkey?” it cannot answer vague questions, even though these tend to be the questions people care about the most. Answering these questions requires background information about the person who asked. Good advice is tailored to the particulars of the asker’s situation.
A natural way to elicit such background is through a dialog in which follow-up questions are asked. If you ask a doctor “Why do I have stomach pain?” they won’t just give you an answer. They will ask clarifying questions: “When did it start?” — “Where is it?” — “Is it constant or intermittent?” And only then they suggest possible causes.
In this set of notes, I propose dialog markets, a mechanism for creating high-quality conversations that resolve vague questions.
A dialog starts when an individual asks a question and pledges a reward that will be used to pay for helpful contributions. Humans and machines then contribute follow-up questions and other responses. The core problem that the market mechanism needs to solve is how to distribute reward over contributions such that we incentivize conversations that are as valuable as possible for the asker. We can view this as a particular set of questions that we can delegate to the market: “How much should the asker pay for this contribution?” This view gives rise to the grounding problem for dialog markets: incentives will be aligned if answers to this meta-level question are good, but that in turn depends on aligned incentives. I outline some approaches towards addressing this problem.
Finally, I suggest an implementation plan: I discuss how to ensure that early dialogs have sufficiently high quality to get the project off the ground and name a few potential initial applications.
- Strategies for Question-Answering
- Break down big questions into small ones.
- Provide answers given by others to turn production tasks into judgment tasks.
- Externalize thoughts to remove memory constraints
- Externalize thoughts to enable incremental progress within individuals
- Share thoughts to enable incremental progress across individuals
- Communicate thoughts to make them more coherent
- Let machines do some of the work
- Focus attention on topics with high expected payoff
- Three Fictional Perspectives
- The System
- Reward Distribution
- Getting Started
- A PDF version of these notes
- An incomplete web app in Meteor/NodeJS (BSD license)
- Some probabilistic models related to reward distribution (BSD license)
This project has benefited from discussion with and feedback from many people, more than I can name here. For particularly helpful and in-depth advice, I thank Paul Christiano, Owain Evans, and Noah Goodman.
I am also grateful to Leon Bergen, Jamie Brandon, Frederick Callaway, Cameron Freer, Roger Grosse, Robert Hawkins, Frauke Harms, Daniel Hawthorne, Sha Jin, Justine Kao, Benedict Küster, Falk Lieder, John McCoy, Jon Malmaud, Long Ouyang, Desmond Ong, Alexey Radul, John Salvatier, Daniel Selsam, Natalie Schaworonkow, Alexandra Surdina, Michael-Henry Tessler, Tomer Ullman, Alyssa Vance, Mike Webb, Jess Whittlestone, and Jeff Wu.