Table of contents
- The problem: answering vague questions
- Dialogs clarify questions
- Dialogs prompt relevant mental computations
The problem: answering vague questions
There are many questions we ask in the course of our lives. For example, there are personal questions, big and small: Which college should I go to? What kind of career is right for me? What should I do in my free time? How can I find a significant other? What mobile phone should I buy? What movie should I watch next? Why do I have stomach pain? Who should I vote for? There are questions that come up in the course of our work: How can I get better at writing? What programming language should I use? What should I work on next? There are big questions about humanity and its place: How did the universe come to be? What are the key scientific and technological developments of our time? When will AI surpass human abilities in most fields? Are humans causing global warming? And there are questions we idly wonder about: Which team is going to win the world cup? What is the best game for the SNES?
Most of these questions are vague: that is, there is relevant information about what the question means that is not explicitly expressed. This doesn’t mean that some ways of answering these questions aren’t much better than others—it just means that the best ways of answering will take into account context.
Dialogs clarify questions
Most of the questions mentioned above are difficult to answer on their own. 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? Is it getting worse or better? Tell me what your pain feels like. Is it dull or sharp? Is it localized or diffuse? How severe is it? Is there anything that makes it better, or worse? Do you have any other symptoms? Weight loss? Vomiting? Fever or chills? And then they suggest possible causes.
So, an important part of answering questions is figuring out what exactly the question means, and what additional information there is that could help answer the question. Typically, this happens through a dialog in which clarifying questions are asked. In this project, we will think about how to set up a system that produces dialogs that lead to high-quality answers. Before we make plans for such a system, let’s first think about another important role that dialogs can play. If you’re in a rush, you may want to jump ahead to Section III, which gives a taste of the overall proposal.
Dialogs prompt relevant mental computations
When someone asks you a question, you think for a bit, then answer. Note how little time passes between question and answer, in particular for real-time dialog. Only very few steps of mental computation can take place between question and answer; and yet, you can answer complex questions like “Do you want to go on a date with me?” in a second or two. Think about all the considerations that would go into answering a question like this if you had unlimited time. What will my future look like if I say yes? What if I say no?
And yet, you can produce an answer for every question in the allotted time, even if it is just a “I don’t know”. Most of the computations used must already have taken place, in your head, in someone else’s, or over the course of human evolution, and you must have cached some of them for future reuse. When we talk about “intuition”, maybe we’re talking about what such cached computations feel like from the inside.
This reuse of cached computations is likely approximate. In Thinking, Fast and Slow, Kahneman discusses such approximations: “If a satisfactory answer to a hard question is not found quickly, System 1 will find a related question that is easier and will answer it.” To give a few examples: “What are the best careers for making a lot of money?” becomes “What careers have I come to associate with wealth?”; “How happy are you with your life these days?” becomes “What is my mood right now?”; and “How frequently does this happen?” becomes “How easily do instances of this come to mind?” (via Kaj Sotala).
Within a conversation, time and resource constraints are similar for questions that require few computational steps to answer and questions requiring many steps. We should expect that, all else being equal, the former will in general get more accurate answers unless pretty much all of the necessary computational steps for the big questions have been cached at some point in the past. More commonly, there will be approximate reuse of computation for at least some steps, so we should expect higher error for big questions.
The same reasoning applies to actions: When we make decisions in real-time, very few uncached computation steps are feasible. For decision quality to be high, we therefore need good cached computations. The source of such computations can be previous computation online (when we made similar decisions), thinking offline (reflection), and external information sources (e.g., things we read on the Internet, or discussions with friends). Consider emergency doctors and martial arts practitioners: almost instantly, they make decisions that you would have no chance of getting right, and what allows them to do this is that they invested years into building the right kinds of cached computations.
Let’s summarize: questions induce computations, and computations partially reuse stored computations. This reuse is approximate when no exact match can be found. If we are interested in the answer to a big question—“What do you want to do with your life?”, say—then directly asking may not lead to good answers; a good answer would take a lot of computation, and there is generally little time available to produce an answer, so if the person being asked has not performed the necessary subcomputations previously, they will have to produce an answer using their best-matching stored subcomputations, which can lead to low-quality answers.
By asking the right kinds of questions in a dialog, we might be able to induce computations ahead of time that lead to better answers and decisions when it counts.