I read this with interest and can't help but contrast it with the research approach I am more accustomed to, and which is perhaps more common in soft sciences/humanities. Because many of us use AI for non-scientific, non-empirical research, and are each discovering that it is both an art and a science.
My honors thesis adviser (US-Soviet relations) had a post-it on his monitor said "What is the argument?" I research w GPT over multiple turns and days in an attempt to push it to explore. I find I can do so only insofar as I comprehend its responses in whatever discursive context or topic/domain we're in. It's a kind of co-thinking.
I'm aware that GPT has no perspective, no argument to make, no subjectivity, and no point of view. I on the other hand have interests and am interested. GPT can seem interested, but in a post-subjective or quasi objective way. That is it can write stylistically as if it is interested, but it cannot pursue interests unless they are taken up by me, and then prompted.
This takes the shape of an interesting conversation. One can "feel" the AI has an active interest and has agency in pursuing research, but we know it is only plumbing texts and conjuring responses.
This says something about the discursive competence of AI and also of the cognitive psychology of us users. Discursively, the AI seems able to reflect and reason through domain spaces and to return what seems to be commonly-accepted knowledge. That is, it's a good researcher of stored content. It finds propositions, statements, claims, valid arguments insofar as they are reflected in the literature it is trained on. To us, psychologically, however, this can read as subjective opinion, confident reasoning, comprehensive recapitulation.
In this is a trust issue w AI, insofar as the apparent communication and AI's seeming linguistic competence elicit trust from us users. And this surely factors into the degree to which we regard its responses as "factual," "comprehensive," etc.
But I am still confounded by whether or not there might be some trick to conversational architectures with multi-turn engagements with AI. Might there be some insight into the prompt structure, or stylistic expression (requests, instructions, commands, formal, informal, empathic...) such that a "false interest" or "post subjective interestedness" might be constructed that can "push" the AI to explore in depth, breadth, novelty, contradiction, by analogy, etc.
For example, four philosophical concepts common to western thinking are: identity, similarity, negation, analogy. Might prompt expressions be possible that serve as navigational coordinates almost, or directions, for use by the LLM in "perusing" discursive spaces (researching within a domain)?
A different kind of argument, a post-subjective kind of reasoning, a different way of taking up the user's interest but nonetheless mirroring it successfully enough that users experience the effect of being engaged in mutually-interested interactions?
I read this with interest and can't help but contrast it with the research approach I am more accustomed to, and which is perhaps more common in soft sciences/humanities. Because many of us use AI for non-scientific, non-empirical research, and are each discovering that it is both an art and a science.
My honors thesis adviser (US-Soviet relations) had a post-it on his monitor said "What is the argument?" I research w GPT over multiple turns and days in an attempt to push it to explore. I find I can do so only insofar as I comprehend its responses in whatever discursive context or topic/domain we're in. It's a kind of co-thinking.
I'm aware that GPT has no perspective, no argument to make, no subjectivity, and no point of view. I on the other hand have interests and am interested. GPT can seem interested, but in a post-subjective or quasi objective way. That is it can write stylistically as if it is interested, but it cannot pursue interests unless they are taken up by me, and then prompted.
This takes the shape of an interesting conversation. One can "feel" the AI has an active interest and has agency in pursuing research, but we know it is only plumbing texts and conjuring responses.
This says something about the discursive competence of AI and also of the cognitive psychology of us users. Discursively, the AI seems able to reflect and reason through domain spaces and to return what seems to be commonly-accepted knowledge. That is, it's a good researcher of stored content. It finds propositions, statements, claims, valid arguments insofar as they are reflected in the literature it is trained on. To us, psychologically, however, this can read as subjective opinion, confident reasoning, comprehensive recapitulation.
In this is a trust issue w AI, insofar as the apparent communication and AI's seeming linguistic competence elicit trust from us users. And this surely factors into the degree to which we regard its responses as "factual," "comprehensive," etc.
But I am still confounded by whether or not there might be some trick to conversational architectures with multi-turn engagements with AI. Might there be some insight into the prompt structure, or stylistic expression (requests, instructions, commands, formal, informal, empathic...) such that a "false interest" or "post subjective interestedness" might be constructed that can "push" the AI to explore in depth, breadth, novelty, contradiction, by analogy, etc.
For example, four philosophical concepts common to western thinking are: identity, similarity, negation, analogy. Might prompt expressions be possible that serve as navigational coordinates almost, or directions, for use by the LLM in "perusing" discursive spaces (researching within a domain)?
A different kind of argument, a post-subjective kind of reasoning, a different way of taking up the user's interest but nonetheless mirroring it successfully enough that users experience the effect of being engaged in mutually-interested interactions?