Keith Douglas
Recap and Response
Alex and Steve answered last month’s column. Thanks to them and any readers who did not write. Alex tells us that AI understanding of psychology is hard. This is correct, and likely involves tacit knowledge; in fact this was the basis of the Dreyfusian critique way back when. Steve has responded by giving us something else to read. I know of the book he talks about; it is cited all the time in the literature and the summary of its contents do reflect what it seems to be about. But I can be slightly embarrassed here – I thought of it when I wrote the most complicated of the sections last time. However, I have yet to read it. If I can easily get a copy soon I might put it in my summer reading. There’s an interesting metaphysical lesson in the study of temperature as well that I’ll address here rather than later (in the detailed discussion). This is that the world really is “dappled” – but not quite in Nancy Cartwright’s understanding. Aristotle seems to have thought there is nothing to say in a place where we can find another kind of problem. This is, that somehow variations in souls, animals, and so on just are: they are not a result of complexity of microscopic parts and surrounding conditions like the modern understanding seems to be. Bunge writes somewhere that individuality is only possible for the big, or something like that. Note however that this betrays in a way his own theory of properties, which allows properties to be relational. (Exercise: use previous columns to find out why I make this remark.) It also brings up ancient theories of mixture – the topic that I was first exposed to in the history of natural philosophy. Studying what we say about that is an exercise much like this and last month’s column – except much harder. I leave that for another day; I’d have to get a contemporary physical chemistry book to feel comfortable confronting modern theories of solutions. Hint: If you want to do this yourselves, you will have to learn that solubility comes in degrees – a fact that superficially, at least, has escaped contemporary studies of dispositional properties.
Detailed Discussion of the Questions from Last Time
In this main part of the column, I will go through many of the times I used a question mark last time and riff a bit on the topic associated:
“Are these appropriate idealizations?” Well, people use them to make very real predictions in practical contexts, so in that way they are appropriate (my father said once they occasionally had to use the ideal gas law in the commercial lab, but never the higher order ones). There are other ways in which they could be (in)appropriate – using the notion of error. This latter consideration is not merely “how far wrong is this” but also a detailed study of when the idealization is expected to increasingly go wrong. We will meet examples of this as we go.
“But how do we know nothing leaked?” I did explore that question last time. But I would like this time to also draw attention to the fact that I appealed to Archimedes. This is epistemologically important, amongst other matters. Why? It illustrates one notion of cumulativity in at least this area of science. If Boyle and company appealed to the work of Archimedes on what happens to metallic crowns in water, they are implicitly assuming that the bag of air is “sufficiently like” and that these areas of inquiry can build up on each other. This seems to be denied to some degree by Cartwright and Dupré and other famous contemporary philosophers who deny the unity of science in the relevant respect. I instead claim that we cannot know “except by trying” that seemingly unrelated fields have nothing to do with each other and to look for mergers of disciplines – which is hard and can lead to dubious ones. Metaphysically this allows us to postulate that the world is such that interdisciplines are possible – again because it seems more fruitful than the contrary.
“the history of how we thought of stuff as being made of little bits of other(?) stuff)” Democritus and the other ancient natural philosophers before Socrates are often thought of as being monists, they held that the world was made of only one sort of stuff. (But is kenon, the void, a stuff? Stay tuned!) Aristotle is ambiguous here – most animals have a soul that is a principle or means of organization – a property, not a thing. When it comes to humans, however, he’s unsure. To my mind it is not merely that there’s a break between humans and other animals being considered, but the fact that the break is in category. Here’s one place you want to investigate your artificial intelligence system. Does it seem to agree with you (or not) about what metaphysical categories items belong to? There’s philosophical debate over some of them in some sense – Whitehead, famously, tries to do without “thing” and use “event” and “process” more.
“What’s that about?” Well, in contemporary understanding, the mean distance between some notion of parts has decreased, rather than losing anything. But this is itself a discovery. In fact, volume is so confusing intuitively that children often think that there’s more stuff in a tall skinny glass than a shorter, squatter glass. AI lesson: does your AI have a notion of conservation that applies here, the matter that children who make the aforementioned mistake, have yet to learn? Some of the image generators don’t seem to have this.
“One is merely(?) the attribution”. The attribution itself is going to be one problem with chatbots, if nothing else. Early versions of ChatGPT 3.5 could very easily get into a “folie a deux” like situation – especially if the user (like a student) was confused. I was able to get it to report that David Hilbert (of mathematics and hotel fame – see earlier columns) and David Hilbert (philosopher of perception) was the same person. This particular instance is fixed, seemingly, but likely ad hocly. One metaphysical principle is that of individuation. What one does the system use to do the basic “same and different” analysis we all do, yet struggle with.
“Similarly at the other end: what happens when pressure goes to zero?” Well, if we naively plug in to (say) PV = k, we get V = k/0. Uh oh. The volume diverges. But I started with an idealization of gas in a container with perfectly rigid sides. This will break down too – oftentimes it is forgotten that the ideal gas law is a model of a system, not just the gas itself. Oftentimes one thinks of idealized pistons, etc. This is didactic, in a way, but remember that historically we are dealing with what was called “the mechanical philosophy”, so mechanical analogies are a matter of course. Another question from my answers to the questions, then: are these analogies? Does your favourite LLM understand them? Hofstadter would argue that analogy making is the basic feature of cognition and so no system can have claim to intelligence without that ability. But does that entail being able to recapitulate brilliant work? Steve’s remarks from last time about students routinely doing experiments and work that would baffle geniuses of the past should be well taken here.
“But what is this temperature?” Temperature of a gas was the subject, but yet we want to also attribute temperature to liquids, solids, and even places where there is very little large scale baryonic matter (matter with mass – this is not a redundancy in modern contexts; realizing that photons have no mass is an interesting discussion – see next time). It is very hard “knitting” all of these notions together; I personally do not know the state of the art when it comes to solids, for example. This leads to the reducibility question. Temperature is often said to be identical in a gas to its mean molecular kinetic energy. However, if temperature in a liquid or solid is something else (and remember sublimation – liquids are “optional”) how does this work “across the gap”? Ask an LLM to explain to you what happens when a liquid boils (Boyles? – would an LLM appreciate that pun?) or a gas condenses. Even the latter is puzzling to us sometimes. My parents when I was very small were concerned about humidity in the basement and eventually got a dehumidifier. But before then they measured the humidity. I asked my father how there could be more than 100% humidity as the dial shows the potential of. I was told this corresponds to being sufficiently waterlogged in the air that it starts to puddle on surfaces. He asked me to remember the botanical gardens we had seen in another part of the city; they would (in the jungle greenhouse) have more than 100% humidity. I tell the story because – again – go through all of that and see what sorts of knowledge are understood in all of it. Even my childhood question shows I had some grasp of how indicators and instruments work. Does ChatGPT?
“in what way justified?” Bunge, who I mentioned, thinks they can be justified by appeals to underlying laws, however imperfectly understood. However – there’s no law of human psychology (?) that justifies the convention that stop signs have a given shape. But there is arguably one that explains why they are red. Yet this seems very different from the convention – equal valence electrons (perhaps) counts- that puts calcium below magnesium in the usual periodic table. Can you ask a question to an LLM that would allow it to show the distinction (if it is one) that I am alluding to? Hint: science vs. technology. If that’s the correct basis, does the AI you are working with share it? Does it matter if the AI itself is for scientific or technological purposes? If it is to lead to “general intelligence”, does it have to grasp that distinction that is (admittedly wrongly in my view) itself contentious in some circles.
“What provoked scientists to complexify that neatly beautiful equation?” I will leave that one to the historians. I don’t actually know why van der Waals did his thing. I don’t even know if he explored the question theoretically (only) as it was in my classes, or whether he conducted experiments, or both … Later of course people have done enough work in both directions.
“What tells us what to use? How do we know we have the model right that we are complexifying? What happened when a model was completely rejected (e.g., that of Aristotle and Ptolemy by Galileo, Kepler, et al)?” I don’t know. This epistemic humility (if I can paradoxically call it that) is another area where AIs must learn – Socrates’ lesson, if Plato can be believed. However, I add one further detail. It has been claimed (I believe by Nobel Laureate S. Weinberg), at least in physics, that since the time of Kepler and Galileo there has been no revolution of the kind in question. Kuhnian changes of view are oversold (in fact,”logically impossible” taken in the strongest way) in my view, and it is unclear what sort of revolutions of thought occur as children mature and as an individual comes in general to learn. However, it is clear that we do individually and collectively “change our minds” and so – how? Does ChatGPT v. GreatThings need to have a mechanism to do this explicitly? Arguably yes. This is where the so-called belief dynamics literature (e.g., to pick two of many names, that of Gärdenfors and my teacher, Arlò-Costa) comes in. Note well, however, that this requires symbolic approaches to AI again to work – or does it? Arlò-Costa’s work (and that of his students) consists often of bringing two approaches to the same problem that seem so different and showing how they relate.
Unfortunately he died a few years ago, tragically young, and so we have a sad legacy to continue. My column continues some of these “joins” and I’d like to think some of the weird connections I try to draw are of the kind he encouraged (even if less formal – one way to join systems of thought is mathematics, which he uses but I do not, at least here). Oddly, this was the same lesson – in completely different areas – that the other great Argentinian I studied with (Bunge) also taught. Ask your favourite LLM what those two have in common. If it answers, “cited by Keith Douglas”, then we’re onto something. Or nothing?
For Next Time
Nothing! No, actually, not nothing. Instead I will write about null objects, null individuals, the number zero, the empty string, a special “empty work product” I am advocating for at work lately, the empty set, vacuums, and so on.