Leiter Reports: A Philosophy Blog

News and views about philosophy, the academic profession, academic freedom, intellectual culture, and other topics. The world’s most popular philosophy blog, since 2003.

  1. Jason Stanley's avatar
  2. Daniel Greco's avatar
  3. Nobody's avatar
  4. Roger of Invisible America's avatar
  5. Santa Monica's avatar
  6. Optimistic about LLM's avatar
  7. Ben M-Y's avatar

Philosopher Daniel Greco is training AI systems to replace him

Brian Leiter

He explains at CHE. An excerpt:

[O]ne of my main takeaways from gig work [for AI companies] over the last few months is just how hard it is to catch the most sophisticated frontier LLMs in philosophical blunders. When ChatGPT first emerged, a common pastime among my philosopher friends was posting screenshots on social media of its more egregious mistakes; prompts involving logical complexity would tend to produce answers that were both confident and comically incoherent. Now, finding errors in the responses of frontier models takes hours of expert paid labor, and the errors you do find — which typically require misleading prompts — are much less blatant than they used to be.

More striking, though, is what happens when you stop trying to trip the models up. When I straightforwardly prompt a frontier LLM with a technically sophisticated and deeply unobvious question — “What challenges are there to simultaneously holding Ramsey-style analytic functionalism about mental categories and endorsing a theistic argument from design?” — the model typically sees what I’m getting at and answers better than I’d expect from my graduate students, of whom I think very highly. I picked that example because it involves two ideas that are not typically discussed in the same context, and so you know the model isn’t just straightforwardly summarizing information from its training data: It has to reason [sic] its way to a novel answer. So when I ask how we should feel about LLMs being good and quickly getting better at philosophical reasoning, I’m talking about the situation as it stands today, not some hypothetical future.

[I added the ‘sic’ since LLMs don’t “reason,” they mimic reasoning, and apparently they’re getting very good at it!]

Professor Greco goes on to argue that AI doing philosophy is compatible with both the intrinsic and instrumental value of philosophy.

Please read the whole piece before commenting!

(Thanks to Philip Bold for the pointer.)

, ,

Leave a Reply

Your email address will not be published. Required fields are marked *

8 responses to “Philosopher Daniel Greco is training AI systems to replace him”

  1. Adjunct

    Having read the whole article, I will just say that my impression of LLMs’ ability to do philosophy is very different. (I typically test only Claude, and only the freely accessible version, so it’s possible that there’s some super-powered AI lurking behind a paywall. But I doubt it.) I do agree about this much: AI does often give first answers that seem reasonable. But when you poke at those answers they still (as of this week, which was the last time I tested it) fall apart: AI gives a *bad vibe* about certain things, but still fails to construct a valid argument. (And when it *does* allude to valid arguments, the arguments come from elsewhere.) I don’t know how to reconcile my data with Greco’s, though I am tempted to say that (since he’s smarter than me) I must be missing something.

    Reply
  2. Anonymous_LLM_User

    I use a paid version of a frontier LLM, typically maximizing the thinking time. My impression is that it’s strong enough to give you an overview of an issue and see tensions between ideas. It isn’t strong enough to produce new research from scratch. You can’t just prompt it with “give me an outline for a high-quality epistemology paper” and expect to get much. (That would be ethically problematic output to use, of course, even if it were good; but in fact, the output is not good.) Still, if I asked it questions like Greco’s, I’d get a competent response.

    I do wonder about the future of logic. ChatGPT is getting very, very good at math. It recently disproved an 80-year-old conjecture in discrete geometry. (Of course, it may be only a matter of time before “give me an outline for a high-quality epistemology paper” does yield such an outline, and at that point I’d wonder about the future of the rest of philosophy too.)

    Reply
    1. Daniel Greco

      For what it’s worth, I agree. I have, with some trepidation, given it a bunch of my papers and asked it to write a new paper in the same style. Maybe this is just ego speaking but I don’t think it did a very good job!

      Anthropic recently released a document I read with interest (linked below) on how they’re using Claude in their own internal operations, which included the following quote: “An area of human comparative advantage, for now, is research taste and judgment, including choosing which problems matter, which results to trust, and when an approach is a dead end.” For now, that matches my sense of where it still falls short in philosophy.

      https://www.anthropic.com/institute/recursive-self-improvement

      Reply
  3. Ben M-Y

    One thing I appreciate about Greco’s piece is the very helpful distinction between instrumental and intrinsic value, which he lays out clearly in the beginning. I also really appreciate his use of the John Henry folktale (see also here:https://www.youtube.com/watch?v=1Qoar03kU0k). But I don’t think he quite sticks the landing when he applies the distinction and folktale to AI (or chess).

    Greco, correctly in my view, says that one moral of the John Henry tale is that technological change can bring about the loss of excellence and the ability to take pride in it. What is lost is of intrinsic value in these cases. But when he goes on to apply this lesson to chess and philosophy, Greco seems to slide to claims about instrumental value and to miss-apply the lesson about intrinsic value.

    Let’s begin with what I take to be the central lesson from the tale of John Henry. When the steam drill replaces John Henry and those like him, the loss of intrinsic value comes from the fact that the industry will no longer call for and cultivate human beings with the excellences he embodied and took pride in. Of course, this comes with trade-offs. Those who wish to ride the rails have more access to the benefits of doing so, and those who wish to turn a profit will reap many rewards from automation. But these are instrumental benefits, and they are benefits for a different group of people. The folktale is powerful, in part, because it attempts to draw our attention to a benefit that is often lost on us, namely, the intrinsic value of skilled labor for the laborers themselves. This is a central tension in the tale, dramatizing a part of the battle between labor and ownership that easily gets lost when we focus just on the broader market access provided by technological advancement.

    How does this apply to Greco’s argument? Consider the case of chess. Here Greco shifts from thinking about the value for the expert to thinking about the value for the novice. He admits that there is intrinsic value in learning to play chess and in playing chess well. But then he goes on to say that this was not lost when people began to be able to turn to computers to teach them to play better, as opposed to excellent chess players. Notice, though, that this is to shift from a case of intrinsic value for the grandmaster teacher to instrumental value for the novice learner. It is to slide from an analogy with John Henry to a missing of the central lesson. The novice player may reap rewards from computer teachers, but the excellent player will lose the opportunity to teach others as she used to, and this will also threaten to undermine the excellences associated with doing so. The world may end up poorer because it has fewer excellent teachers of chess and fewer opportunities for humans with the ability to do so to exercise that and take pride in it.

    The same problems occur with respect to Greco’s discussion of philosophy. Granting that there is intrinsic value to doing philosophy well, Greco invites us to “imagine a world where people who want to think hard about philosophical questions form salons where they read and discuss together, except no there’s no professor at the front of the room charging tuition.” Even supposing he is right that this is a world in which more people can do philosophy well, it leaves totally unaccounted for the loss of intrinsic value for those who excel at and teach philosophy of the practice of doing so. Greco mentions that he likes his job and wants to keep it. But that’s beside the point. The question, insofar as we are sticking with the John Henry analogy, is not whether a professional philosopher should continue to be paid for his work. That issue concerns the instrumental value of doing philosophy. Rather, the question is about the intrinsic value of doing so well and taking pride in it. That need not depend on it being one’s job at all—as Socrates pointed out long ago.

    Greco says that “the important question for most people isn’t whether current members of the academic philosophy guild get to keep their jobs. It’s whether the world is made better or worse by AI getting good at philosophy.” True! But the distinction between instrumental and intrinsic value, and the moral of the tale of John Henry, is that we cannot forget that the world can be made worse by technological change that threatens the development and maintenance of expertise that one is justified in taking pride in. This may, in the final analysis, be outweighed by the instrumental goods in question. But it cannot be reduced to them, nor will it show up in our cost-benefit calculations if we ignore the perspective of those with the ability and opportunity to exercise the excellence(s) in question.

    Reply
  4. Optimistic about LLM

    Some users seem to have conflicting impressions about the (absolute level of) abilities of current iterations of the LLMs. But it seems to me beyond serious doubt that the current iterations of LLMs are astronomically better relative to previous iterations even just a year ago. And I see no reason to doubt that it would continue to improve in this way, especially considering how Claude is moving towards “recursive self-improvement” (it designs and trains its own successors). That may help in terms of doing the kind of self-guided, original research that some find current iterations unsatisfactory. In any case, even those who are skeptical about the current abilities of the LLMs may well reasonably expect stuff to change in the coming years.

    For whatever it’s worth, my impression with the latest paid model of Claude is strikingly positive: it could easily and elegantly prove everything I need to prove in my paper, and could spot errors and room for improvements as competently as a graduate student in philosophy. It also digests exceptionally well research papers published in top tier philosophy journals. And you could easily build a unified “project” by giving it enough “project knowledge” on the magnitude of several books.

    Reply
  5. Santa Monica

    Conditional on the hypothesis that journal editors sometimes desk reject high-quality papers written and commended by world experts on a prominent topic for which the journal editors are not world experts, I would guess that some philosophers would not mind it if editors used such specially trained LLMs to avoid desk rejecting high-quality papers written and commended by world experts on a prominent topic for which the journal editors are not themselves world experts. So, perhaps AI that is good at philosophy could actually be to instrumentally valuable to talented professional philosophers’ even without those philosophers ever using AI.

    Reply
  6. Roger of Invisible America

    Perhaps a useful distinction here is between replacing philosophers and changing the conditions under which philosophical work gets done. LLMs are not philosophers. They do not understand, intend, take responsibility, or care whether what they produce is true. So Leiter’s “sic” seems right if “reasoning” is meant in the ordinary agential sense. But it does not follow that LLMs are merely inert tools. A library, notation system, referee report, or good interlocutor can reorganize inquiry without being an originating philosophical subject. Harold H. Oliver (1930–2011) https://leiterreports.com/2024/05/01/bradley-vs-russell/comment-page-1/#comment-6197, mentioned in an earlier Leiter Reports thread on the metaphysics of relations, is useful here. His basic point was that relations are not external add-ons to already self-contained things. What something is often depends on the relations through which it acts, appears, and is taken up. Applied to thought, that means philosophical agency is not the private emission of a sealed interior ego. It already works through language, inherited problems, institutional norms, books, conversations, and now machines. The real question, then, is not whether AI “does philosophy” in the human sense. Plainly, it does not. The question is what kind of philosophical relation it can enter. Used well, it may function as a stress-test, objection generator, or map of argumentative terrain. Used badly, it becomes a device for producing fluent substitutes for thought, which academia was hardly waiting for AI to invent.

    Reply
  7. Nobody

    Looking at some of the AI-amplified cranks on philpeople, it is hard to agree with the assessment of this article. Maybe they don’t have “the most sophisticated frontier LLMs”?

    Reply

Designed with WordPress