Since the release of ChatGPT in late November of 2022 I have logged hundreds of hours interacting with LLMs. Most of those hours have been with ChatGPT, but about two months ago I started using Claude 3.5 and most of my LLM-time has been with Claude in those months. Many of these sessions have made their way here to New Savanna, where they are tagged with the appropriate label: ChatGPT, Claude.
During that time I seemed to have developed five modes of interacting with an LLM. The boundaries between the modes are fluid. Within a given session, I may move from one mode to another. But there are also sessions involving only one mode. This post is an initial attempt to classify these modes.
Exploratory
This is where I began with ChatGPT back in 2022. I just wanted to see what the engine would do, how it would response to a particular prompt. It’s not clear to me that any session ever remained in exploratory mode the whole way through. Rather, I’d initiate a session in exploratory mode and then switch to one of the other modes as seemed appropriate.
For example, when I approached ChatGPT with the intention of seeing whether or not it would be able to analyze Steven Spielberg’s Jaws using the ideas of Rene Girard, I began that session in exploratory mode. As soon as I saw that it was working, however, I switched into object mode.
There are many sessions where I begin in one of the other modes and stay there for the duration of the session.
Object
In this mode I am treating the LLM as an object whose behavior I am examining. I want to see what it can do, how it behaves. Much of my work with ChatGPT has been in this mode. This work has been fairly systematic, so systematic that I could probably develop protocols that others could use to duplicate the tasks I’d set for the LLM. See, for example, this working paper about stories.
Much of my recent work with Claude has involved having it describe a photograph. Almost all of these sessions are undertaken in object mode. I want to see what Claude does when asked to describe a photo. A few of these sessions became quite long. In those cases I suspect that I have slipped into exploratory mode within object mode.
Query
This is fairly straightforward. I’m looking for information. Thus I’ve had quite a number of sessions with Claude in connection with my inquiry on melancholy. All of those sessions have been in query mode.
Play
I wasn’t expecting this. But when it emerged, I found it delightful. It probably began when I was working on stories with ChatGPT. I believe that first extended example of this is the story I improvised a story about the OpenWHALE, a whaling vessel on a reality line that collides with the Starship Enterprise. The Green Giant Chronicles is another example. Both of these interactions took place over multiple sessions.
The signal characteristic of these sessions is that I laugh, often out-loud and with noticeable body movement. There is certainly and exploratory element here. I want to see how the engine will react to this or that prompt, where I really don’t have any reasonable expectation of what it will do. But I want it to provoke laughter.
This mode is very important, for I believe that it is my experience with this mode that led to the last mode, interlocutor.
Interlocutor
In this mode I am treading the LLM as a dialog partner, we’re working something out together. We’re not equal partners here. This is not like sessions I used to have with David Hays. I definitely take the lead in these sessions. But I’m not simply looking for answers to queries. I have an idea I’ve been thinking about – sometimes for a long time – and I’m using the LLM as a sounding board. Does this idea make sense? I’ve only used this mode with Claude 3.5. I might be able to use it with ChatGPT, but why bother?
I’ve uploaded a number of these sessions and labeled them Claude_collab (for collaborative or collaborator). These are shade into sessions I’ve labeled Claude_partner. I’m not sure of the distinction between the two, and there are sessions with both labels.
No comments:
Post a Comment