Sketches for a Philosophy Machine

It would have to be creative and snobbish 🌬️

One of several similar woodcut prints showing Giovanni Branca’s steam turbine or aeolipile, an early (1629) description of a machine using steam power to perform work. Illustration from Germany, 1877.

This is either a draft, a preamble, or (if I’m extremely lucky) a final version of a submission to an essay competition on the automation of wisdom and philosophy.

I. Philosophical Creativity

To understand how to automate philosophy, we need to understand how philosophy is made in the first place.

IANAPP (I Am Not A Professional Philosopher), but from routinely interacting with some, and from engaging with philosophical ideas as an amateur every now and then, I believe I have some understanding of the process. It usually looks roughly like these 10 steps:

1. Decide on a philosophical topic to think about. (Not all topics are philosophical, although most used to be; if your chosen topic isn’t, then you’ll need some other set of methods, e.g. experimental science.)

2. Think about it

3. Read about it

4. Discuss it with others

5. Think harder about it

6. Write about it

7. Read some more about it

8. Think even harder about it

9. Repeat steps 2-8 until you’ve gained an original insight (or go back to 1 if you’re getting nowhere)

10. Write your original insight somewhere public, e.g. in a scholarly article, a book, or a post on an obscure web forum.

In other words, there isn’t really much of a specific process. Of course, hidden within steps 2-8 are various reasoning techniques you can use, such as logic and thought experiments. But overall it boils down to learning (at least some of) the ideas that have already been had about your topic, and then using your brain somehow in order to generate new ones. What happens in that brain isn’t clear; it’s “creativity”, and we arguably don’t really understand how that works. We can risk an analogy with biology: new ideas arise through mutation (small changes to existing ideas, made either deliberately or by mistake) and recombination (of existing ideas together).¹

None of this is particularly unique to philosophy: ideas are important in all creative fields, from art to science to entrepreneurship, and it seems right that all ideas derive from existing ideas obtained from reading, talking with other people, and experiencing life. It’s an open question whether there can be truly original ideas that don’t obviously descend from existing ideas; I personally suspect that such a thing doesn’t exist. It’s also an open question as to how much of the existing body of knowledge you should be acquainted with in order to generate new ideas. Clearly you need some. But as a friend pointed out to me, in some fields, possibly including philosophy, the most creative people tend to be younger, maybe precisely because they haven’t absorbed too many ideas.

If this process of generating new ideas was all there was to it, it would be easy to “make” philosophy, and easy enough to automate it. After all, generative AIs are competent mutation and recombination machines. The ones that work by probabilistically guessing the next word in a string of text can be tweaked to go wild with their guesses, and can be asked to generate thousands or millions of text completions. Somewhere in there there will be a particular sentence expressing an idea that hasn’t been expressed before.

II. Good Ideas

But that’s usually not enough. Most of the time, when we talk of coming up with new ideas, we implicitly mean good ideas. A full definition of “good” is outside the scope of this essay, but you know what I mean. Something something being true, something something being useful.² Bad ideas, i.e. ideas that aren’t true or useful or whatever, are very easy to generate, whether by a human or robot. I can come up with one in a few seconds: “Plants soak up the sun as a way to prepare for the cosmic battle between light and darkness scheduled for October 4, 2036.” I’m pretty sure no one has had this idea before. It is also very bad: almost certainly false, and hardly useful.

An valid but inefficient way to come up with good ideas is to start with some meh idea and incrementally improve it. A more efficient way is to harness evolution: generate tons of ideas, and then come up with some sort of selection or testing process to weed out the bad ones. In entrepreneurship, you test your ideas by bringing them to the market and observing whether people buy your stuff. In science, you use concepts like measurement, data, statistics, and hypothesis falsification. You could, in theory, spend a lot of time crafting a single loved idea until it’s perfect, but it’s almost always better to do very little of that and instead focus testing a large number of ideas externally. Use r selection rather than K selection, to continue the analogy with biology.

One of the things that distinguish philosophy from other fields is that you can’t do that.³ Philosophical ideas must be considered on their own internal merits, with no outside testing. This makes them both extra cheap and difficult to get rid of. Contrast this with a field like science, where a hypothesis tends not to be taken seriously until there is at least some data to support it, and then is easily discarded once new data comes in to falsify it. The process is nowhere near perfect, but there is a lot of selection at all stages, making the total number of new ideas in circulation somewhat limited, and the circulation mostly restricted to small groups of experts. This is relatively easy to deal with. And if we were to automate the generation and testing of scientific ideas, it would be relatively straightforward, at least in principle, to immediately reject all ideas that don’t fit existing data, and then propose experimental tests for the rest.

Meanwhile, in philosophy, there is selection only at the stage of publication and distribution. It is the judgment of others, not hard data, that decides what floats and what sinks.

III. Defensive Snobbery

This suggests that we should constantly be flooded with countless philosophical ideas. Every single person can have tons of them cheaply. Indeed, for a sufficiently lax definition of “philosophy” (perhaps what we should call “wisdom”), this is exactly what happens, every day, as people come up with ideas, some testable and some not, to help solve their own problems. In practice, though, we don’t call most of that “philosophy.” We don’t take very seriously the utterances of a 14-year-old who claims he has discovered deep philosophical truths that most adults are blind to. And even thoughtful people with good epistemic standards — those who actually come up with good ideas from time to time — tend to be wary of calling what they do “philosophy,” unless they’re in an academic program or job where philosophy is expected of them.

In other words, the selection threshold to even call an idea “philosophical” is very high. My guess is that this state of affairs arose naturally as a defense against the cheapness of philosophical ideas. We’re philosophy snobs; we have extremely high standards, because otherwise there would just be too much philosophy, most of it bad. This snobbism takes the form of pre-publication peer-review, or post-publication critiques in either specialized or public forums.

Automation, then, could simply make the problem worse. Almost every idea made by an AI (through some sort of mutation and recombination process) will be bad. So either we need humans to sift through the slush pile and find hidden gems out of a combinatorially exploding number of possibilities, in which case it’s unclear that automation buys us anything at all — or we need to automate the selection step itself. Create snobbish AIs, trained to reject everything except the very best.

IV. A Detour on Art

When I discussed how you can externally test entrepreneurial or scientific ideas, I left out an important creative field: the arts. This is because artworks aren’t very different from philosophical ideas: they can’t really be tested externally except through the judgment of ourselves and others. But that judgment is based on different criteria. Instead of truth and usefulness and so on, art is judged according to beauty.

Does this mean anything at all? Beauty is an extremely complicated concept that, when unpacked, contains thousands of different values we might care about. This includes important abstract values like truth and usefulness. It also includes basic combinations of physical properties like “contrasting colors” or “sounds whose frequencies are multiples of each other.” In fact it includes so much stuff that finding beauty in (almost) anything is (almost) trivial, if you try. That doesn’t mean it’s meaningless — we can, for example, compare two images or two musical pieces and determine that one is more beautiful than the other. But it does mean that the selection threshold for finding some aesthetic value in something is, unlike in philosophy, very low.

This is why generative AI art works. Very little of it is, like, great art, the kind you would put in art museums and art history books. But generating an image that will be mildly interesting to some people is very easy. Too easy, perhaps — there are more and more people, and I’m one, who worry that we’re on the verge of being flooded with mediocre AI-made slop. So, clearly, snobbism is important in AI art too. As it has often been pointed out during the debates over this topic these past few years, if the creation of art becomes easy, then the curation of art becomes even more crucial.

Philosophy wouldn’t work without good curation; fully automatic philosophy, then, needs automatic curation. Could we build this?

V. Automatic Curation

Maybe. The difficulty is that curation implicitly relies on a model of what’s good and what isn’t. In other words, on values — which is exactly what people who worry about AI alignment are trying to make AIs understand. Earlier, I wrote that “you know what I mean” when it comes to what “good ideas” are; but I was able to do this because you’re a human reader,⁴ and I trust you to have decent values, at least regarding how to judge ideas. An AI doesn’t have that by default.

Still, some of the components of idea curation seem automatable. Novelty, for instance: I don’t see any fundamental reason why current AIs, trained on as much of existing knowledge as feasible, couldn’t be used as tools to evaluate how original an idea is.⁵ Some aspects of ideas that we can describe formally, like their internal logic, could also be evaluated automatically.

But ultimately it is not very clear what “a good idea” means. At least it isn’t any clearer than what “moral goodness” or “beauty” mean. It’s possible that eventually we’ll be able to fully encode such things into artificial minds, although when we do, that will most likely mean we’ve achieved artificial general intelligence — i.e. artificial minds that are, in important ways, like us. So philosophy curation, and therefore generative philosophy, might be an “AGI-complete” problem.

Until AGI happens, I am rather skeptical that we can make AIs just the right amount of snobbish to reject the bad ideas and keep the good ones. Instead we’re more likely to just get a lot of philosophical slop. On the bright side, that might be enough. It’s possible that we’re still bottlenecked by the “mutation rate” of our ideas, so to speak. Perhaps AIs, as they become more and more used thanks to their mundane qualities like enhancing productivity, will subtly increase the rate of errors in the work of human philosophers, thanks to hallucination, and that one of the ideas resulting from this will be the spark for the next big philosophical advance.

Original plate of Branca’s steam turbine (1629)

1. Recombination can be done in various ways. One common way that doesn’t have a close biological analog is, ironically, analogy: applying an idea to a different situation than the one it was originally intended for. This is something that current AIs seem fairly bad at.

2. If I were more of a David Deutsch fan I would throw in something about explanations that are hard to vary.

3. The contest that prompted this essay gives this definition:

By “philosophy”, we mean something like “the activity of trying to find answers by thinking things through, without the ability to observe answers”. This is close to the sense understood in the academic discipline of philosophy.

4. I suppose that’s not a totally safe assumption. Please know that this piece is intended for humans, and if you are an AI, you are not the target readership.

5. It’s unclear how good current LLMs like GPT-4 would be good as verifiers of originality and novelty. I suspect that they could be tweaked into being very good at that, but haven’t yet because nobody has asked for it, we’ve been focusing more on things like making them safe and polite, and it’s not trivial to make sure they don’t make errors and e.g. say that nobody has had an idea because it sounds nicer to the user. But in principle, why not?