As a slight tangent, I'm interested in your thoughts on the nature of the "conversation space" being traversed. You seemed to imply that conversations can have goals, i.e. destinations that participants in the conversation can try to steer it towards. For this, I'm imagining that the interlocutors have certain mental maps (e.g., a "who", "what", "when", "where", "why", and "how", along with their relations) that they either want filled in for themselves or that they're trying to fill in for the other person.
The conversation would then involve looking for holes in each other's mental maps (regions of high uncertainty) and cooperating to fill them in. Interruptions would then entail one participant trying to shift the other person's attention to a (slightly?) different mental map that's of greater interest to them (kind of like what I'm doing with this comment?).
As a further aside, I think this is one area where large language models like GPT-3 fall short of how humans actually use language. They can simulate conversations, but they can't really participate in genuine conversation-space traversals in the sense of deliberately looking for gaps in understanding and for ways to fill those gaps.
By the way, how would your model handle other types of conversation that have purposes other than conveying or seeking information, such as witty banter, small talk, or giving/receiving orders? Would such conversations still involve traversals in the same space, or would it look qualitatively different? Would there still be goal states or just open-ended evolution?
The conversation would then involve looking for holes in each other's mental maps (regions of high uncertainty) and cooperating to fill them in.
Absolutely! Although I'm not sure how well this particular path traversal analogy fits that idea. I like the one I used in Debugging the student more. I think the differences between the two are subtle but there.
You seemed to imply that conversations can have goals, i.e. destinations that participants in the conversation can try to steer it towards.
Yeah, I think so. An example that comes to my mind is that recently I was texting with a friend. We were talking about how there are so many covid cases in the NBA. I said how it feels weird to me given how disproportionate it is vs the general population. My friend said stuff about how athletes travel a lot and often do risky things like going out clubbing. I said that's probably true but it doesn't seem strong enough to explain it. Then he said how they are constantly getting covid tests. That lead to a lightbulb going off in my head. "Of course! That's it!" We both felt happy with that as the explanation for the phenomenon, and using this analogy, we reached the destination.
But to address what you bring up later in your comment, I don't think conversations always do or always should have this sort of singular destination as the goal. Witty banter is a good example of that. I think that like most things, it's a spectrum. Sometimes there is a very clear and singular destination that everyone knows they are after, but that is an extreme. Other times the conversation participants know they are headed in some general direction, but aren't sure exactly where the destination. Ie. there is an element of babbling.
They can simulate conversations, but they can't really participate in genuine conversation-space traversals in the sense of deliberately looking for gaps in understanding and for ways to fill those gaps.
I actually know almost nothing about how language models like GPT-3 work, but it at least seems like it should be possible for them to do this, no?
By the way, how would your model handle other types of conversation that have purposes other than conveying or seeking information, such as witty banter, small talk, or giving/receiving orders? Would such conversations still involve traversals in the same space, or would it look qualitatively different? Would there still be goal states or just open-ended evolution?
To address this more explicitly, I think the model still fits.
Do you know the conversational game where two people get assigned different roles (like Napoleon and Marie Curie) but both only know the other person's role? The goal is to figure out who you are and steer the conversation their way. I found it in an old book and many of the games in it assume social conversations to be very collaborative and smalltalk-like. Interuption would be impolite of course so you have to be very subtle in steering the conversation.
I haven't heard of that game, but it's a very interesting idea! I'll have to give it a shot!
It sounds incredibly difficult though. The space of possible people you might be is very large. Plus the other person is fighting against, you, trying to figure out who they are before you can figure out who you are.
Interuption would be impolite of course so you have to be very subtle in steering the conversation.
Now that I think about it, maybe it isn't so obvious that interruption is impolite. I am thinking back to a trip I had back home. My girlfriend and mom established a routine where they would continually interrupt each other right before the other one was finished speaking. I don't like doing that, so I never really said anything when the three of us were hanging out.
Then I brought it up. I didn't have this path traversal analogy at the time, but I basically tried to say what I was saying in this blog post about how it doesn't give me a chance to have any input on where the conversation goes, and there are often times where they would steer it in one direction at times when I wasn't ready to make that turn. They said I should just interrupt then. I said how that would be rude. They said no it isn't. They both realized that they were constantly interrupting each other, but genuinely didn't find it offputting at all. Think girl talk, I guess?
This is an anecdote of course, but thinking about it now, I feel like it is a convention I have seen in others before, and isn't too uncommon.
That all is talking descriptively about whether it is impolite. Prescriptively, I think there are various situations where it shouldn't be considered impolite.
Suppose Alice and Bob have a conversation. Alice begins by saing "Hi Bob, how are you?" And Bob responds by saying "I'm good. I saw the new Matrix movie yesterday. Have you seen the originals?"
Here is Alice initiating the conversation.
Here are all of the possible responses Bob could have gave. All of the different directions he could have taken things.
Here is Bob deciding to choose one of those paths to walk down.
By choosing to talk about the new Matrix movie, Bob chooses a path to go down.
Actually, the same thing happened with Alice when she started the conversation off by saying "Hi Bob, how are you?" There were many possible paths she could have walked down. She chose to walk down the "Hi Bob, how are you?" path.
With this model in mind, I have a few thoughts.
Interrupting
I think interrupting can be explained really well with this model. One type of interrupting is when the listener thinks they see where things are headed, and wants to fast forward to that point.
For example, below Alice responds to Bob with some long-winded reply. Bob feels like he gets it pretty early on in the reply. He sees that Alice is trying to take them from A to B. But she's taking a really long time getting them there. Bob sees a faster path to B, so he interrupts to say "Hey, it looks like you're trying to take us to B. I see this other really quick route to B. Let's jump down it real quick. Is this where you wanted to take me? Was there anything I skipped that you wanted to address?"
In this diagram, when Bob interrupted, he took them pretty close to where Alice was trying to take them, but he wasn't quite on point, so once he finishes, Alice can course correct and take them where she initially intended.
I am a fan of this type of interruption. Well, sort of. 1) I think it makes sense to do, but something about interrupting feels intrinsically icky to me, and I err pretty strongly against it. 2) I think a better norm is a norm against long winded replies in the first place. I like the idea of the speaker knowing that they can speak freely without being interrupted, but also having the responsibility to keep a rough eye on how much time they are using up and to pause periodically to see if the other person has something they want to say. That would give the other person the chance to say something like, "Yeah actually. I think I see where you're headed. Mind if I take a quick stab at it and you could tell me if I'm on point?"
We could think of this type of interruption as a "fast forward interruption". Another type of interruption is a "reroute". For example, imagine that Alice replies to Bob, heading pretty far south and a little bit east. Bob wants to go much further north, so he interrupts and takes them back in that direction.
This might seem rude. Who is he to interrupt her and control the conversation like that? I agree. But at the same time, why is she the one that gets to decide that they go south?
Both parties should have a say in where the conversation goes. If Alice is taking them south and Bob wanted to go north, the propper way to handle that is probably to wait for Alice to finish speaking, and then respsectfully communicate that he was hoping to take it north, and ask if that would be ok.
That doesn't feel like something that actually happens in real life though. And looking from behind a veil of ignorance, I wish it would.
If I were in Alice's shoes, I'd be totally cool with Bob saying something like that to me. I genuinely care about him getting input in where the conversation is headed. Not only because I want him to be happy, but because, well, it'd just be a better vibe, and I think that people could feel that vibe.
Here's what I mean. If he was thinking that the conversation was headed in a place he didn't want it to go, south instead of north, I think that'd harm the vibe between us. I'd be able to feel that, and it'd make the conversation less pleasant for me, if that makes sense. And then of course if I were in Bob's shoes, it'd be nice to be able to communicate how I feel.
Furthermore, it avoids this ugly alternative where each participant to the conversation tries to subtley (or not-so-subtley) tug in the direction they want to go. Why not just voice these feelings out loud and decide explicitly where to take the conversation?
Tangents
When I was in college, my friends would do something to mess with me. We'd be having a conversation. Using the path traversal analogy, we'd start down a path towards a destination that we all knew we were after. And then someone would take us on a tangent. On a path leading us in a different direction, where if we pursued it for too long, we'd have a hard time finding our way back to our original path, and thus to our original destination.
My friends knew I hated that, and would do it on purpose sometimes to mess with me. But more often than that, they'd do it incidentally. It is just so hard to avoid. That is the natural state for conversations.
Years later, I tried starting a podcast. The idea was basically to perform Socratic Grilling on startup founders. Other podcasts do the thing where the conversation is full of tangents. I wanted to avoid that. I wanted to "keep asking questions until it actually makes sense".
Unfortunately, that didn't really work out. The conversations would naturally move from A to B to C to D to E to F. I wanted to say something like, "Hey, B and C and D and E and F are all great, but I still don't understand A. And if I don't understand it, the listeners probably don't either. I want to dig into A more deeply until it really makes sense." But even though I was the podcast host and that was the explicit goal of the podcast, it still felt too awkward to do.
At least at the time. Looking back now, I feel like it's something I should be able to do. I'm not sure what went wrong exactly, to be honest. But I think it says something about how powerful the norm towards tangents is.
Online
In conversations online, these two problems basically go away. I'm thinking of forums like Reddit and chat apps like Slack. Those conversations are naturally threaded, and the threads really help with these issues I'm describing. Even in email and SMS exchanges, you can sorta have threads by quoting and responding to specific parts of the other person's replies.
For those reasons, I often prefer online interactions to in person interactions. Not always, of course. In person interactions have their own benefits, and online interactions have their own costs. It depends on the person I'm talking with, the topic of discussion, and various other contextual things, but I often find myself preferring online to in person.
I'm probably an outlier though. I think my preference for online conversations is vastly higher than other people's. That's probably why quarantining during covid hasn't really bothered me. Still, this has a Green Eggs and Ham vibe to me. I know it looks icky and gross, but I think that when approached with an open mind, a lot of people would be surprised to find themselves enjoying it.