OpenAI now offers a speech interface to GPT. I’m told it feels like normal conversation. One is talking to a text corpus, bluring the line between “dead” texts and “live” conversation. It prompts me to wonder about Plato’s objections to writing and preference for dialogue in works like the Phaedrus, particularly in the story of Thamus & Theuth and subsequent discussion.
Writing talks back
With a speech interface to GPT, writing no longer has “this strange feature, which makes it truly like painting. . . . if you ask them something, they preserve a quite solemn silence” that Plato’s Socrates asserts (Phaedrus 275d5, transl. Rowe). Works will no longer be vulnerable to Socrates’ objection (if they ever were) that “when once it is written, every composition trundles about everywhere in the same way, in the presence both of those who know about the subject and of those who have nothing at all to do with it …” (ibid. 275e).
ChatGPT certainly won’t be the enlightened philosopher of Plato’s ideal case as it lacks the wisdom of one who recalls the Forms. The response of a large language model (LLM) will incorporate material beyond the single document queried, depending on the scope of the text being searched and the model training. If the corpus is, say, all Plato’s texts, that might approximate to how Plato’s Socrates would respond to questions in a dialogue. If the corpus includes subsequent commentary, the “speaker” won’t just be Plato. It could be as good as a pedestrian professor leading students through Plato.
As an example of a tailored LLM, the Technology Policy Institute (TPI), a DC think tank, has created ChatTPI, an LLM trained on all their papers. This model drafted a filing to the FCC responding to a new rulemaking that the president of TPI said “isn’t good enough to submit, but it’s close” in a WSJ op-ed (see TPI’s human and AI filings).
Last December I speculated that LLMs could enable conversations with, and between, ogregores. ChatTPI validated my “If I can think of something, someone’s already done it” rule. Having recently read the Phaedrus again, I’d love to see LLM fine-tuned on Plato’s dialogues so that I could converse with his Socrates. For added entertainment: imagine a Plato bot debating a Nietzsche bot.
Limitations of Generative Plato Transformers
Paul Diduch pointed out to me (personal communication) that a core problem with training an LLM on Plato's corpus is that his writing requires artful reading. It demands more from the reader than merely absorbing surface-level information. As Paul puts it, one must read and interpret the utterances in Plato's dialogues in the light of the drama and the characters' psycho-dynamics. Plato designed his dialogues to minimize the problems of a written text that he highlighted in the Phaedrus, and to mimic in-person conversation as much as possible. He used the interplay of argument and action to create a depth that can be interrogated in a way that resembles interlocution.
Certainly, the best form of education for most people is 1-1 tutoring by a caring and responsive guide. I strongly doubt that current bot technology can do this. There may well be something uniquely human in a good teacher’s ability to understand their interlocutor's needs that will only be imperfectly imitated (at best) by bot tech.
However, one should compare like with like. Very few people get one-to-one tutoring from other humans, and very few indeed have a guide that tailors responses to their needs. How many teachers apply a sophisticated theory of mind and high self-awareness to every student (or any student)? Such individuals are scarce. Sages like Socrates and geniuses like Plato only come along once in a millennium. Bots will be imperfect, but most people’s ability to understand others’ needs are also flawed.
Good enough is good enough
Consider the critique that AI radiologists aren’t as quite good as humans. Orly Lobel reports in The Equality Machine (2022) that an AI radiology screening algorithm performs slightly worse than — or the same as — two highly trained human radiologists. In most parts of the world there are hardly any trained radiologists; an AI operating as well as a single competent scan-reader could revolutionize care for billions.
Even if a ChatPlato bot never reached the level of a skilled human interpreter, let alone a writer of genius like Plato, it could help laypeople like me engage with the material. It couldn’t replace careful reading and discussion with a good teacher but could supplement learning, similar to podcasts and lectures..
Generative Plato Transformers in the classroom
Some educators may feel threatened by bots replacing them, assuming (optimistically) that bots could be as good or better than the average professor in leading students into a philosophical question or text. (I certainly suspect a fine-tuned LLM could replace me in the T&M seminar, especially teaching over Zoom.) Even a mediocre simulation could be a net advantage for students as a whole, especially in areas lacking competent human teachers.
It would be a mixed blessing, of course. It would harm students who lose access to a good human teacher, and it would be a loss for culture at large if there are fewer cross-subsidies from teaching to support scholarship.
A question to consider right now is how to track the performance of bots vs. humans in educational settings. Lots of methodological challenges, from hidden assumptions and begging the question in experimental design to what’s observable, and whether observables reveal outcomes.
I don’t think bots will replace human teachers and classrooms, any more than TV did (as Neil Postman expected would happen). LLMs do, however, raise the question of what the role of human educators will—and should—be in a few years. They will certainly continue to tutor elites, just like elites still go to live theater even in an age of on-demand streaming. Another possibility is that humans will “teach the teachers,” i.e., fine-tune the bots that directly teach the mass of students.
Bottom Line
Plato’s dialogues are literary works, not conversation transcripts. It’s unlikely that one can learn much from them about what it would be like to converse with Socrates. In the same way, plays are artful constructs not a “realistic” record of events. (Whatever an objective record of an event might be; even real-time video footage is taken from a particular perspective and can be open to multiple interpretations.) It’s implausible that one could write new plays by Aristophanes or chat to Oedipus based on Sophocles’ Theban plays by training an AI, though this week’s AI sensation of (the deceased) Johnny Cash covering a Taylor Swift song might give one pause. I would love to have a go at training a Shakespearean Fool, though…
Coda
After drafting this post, Paul Diduch shared a Jeff Holmes tweet with me: “Try my chat app” becomes the new “check out my podcast.”
Postscript
Researching this post I ran across researchers using the Socratic method as an approach to prompt engineering:
- Runzhe Yang and Karthik Narasimhan have proposed SocraticAI, a new method for facilitating self-discovery and creative problem solving using LLMs.
- A Constantin Brîncoveanu blog post explores the Socratic method, System 1 and System 2 thinking, and the Tree of Thoughts prompting technique
- An arXiv paper by Edward Chang presents a systematic approach to using the Socratic method in developing prompt templates that effectively interact with large language models, including GPT3
No comments:
Post a Comment