Thought Strewn All Around Us

Fluidity Forum '23 Talk and Reflections

Last month, I spoke at the inaugural Fluidity Forum. I’m going to share my talk a bit further down, but first I’d like to talk about Fluidity as a whole. One way to encapsulate Fluidity Forum is that the first presentation brought the house down with this meme:

If you get the joke here, you’d fit right in. If you don’t — then the elevator pitch is that this is a spot for people who like to think about thinking. There’s a whole social scene around this, certainly, and knowing the canon will probably help you appreciate some of the nuances. But folks at Fluidity are unusually kind and willing to explain, so all you really need is curiosity and a willing to interrogate your lived experience more strongly than the norm. In fact, one of my favorite talks was about the Qualia Diversity Project, and all you need to contribute to that frontier of research is literally having been alive. (Even if you didn’t go to the talk, you can still help the cause by taking the survey, and it can be very interesting to see which things get divided into categories that you assumed were universal experiences.)

A lot of the fun was the conversations that came after, which is the mark of a good conference. And it had a rule that everyone who came had to bring something (a presentation, exhibit, food - no restrictions aside from “not nothing”) which anecdotally seemed to filter out wet rags and got us an interesting and assorted gang of people. If you’re interested in going next year, it’s worth subscribing to the newsletter or reaching out with any questions.

For my contribution, I went with the talk “Thought Strewn All Around Us”, which linked together some analogies I’ve used before to highlight a central theme: the mental shift of thinking of facts/memories/knowledge not as discrete things unto themselves but pointers at patterns sustained in environments through time. I was most explicit about this in Memories are environmental indices, but it’s lurking behind a lot of what I post, and it was worth calling out as a whole.

From the beginning, I knew I wanted this to be an extemporaneous speech with just a rough outline and no associated PowerPoint. I’m metaphor-building and not pushing individual facts around, and the ability to be expressive and spontaneous was a huge benefit for that. However, as every standup comedian knows, the secret to having a good spontaneous speech is to have a gigantic library of soundbites already in the top of your mind and to save the spontenaity mostly for stitching them together. I planned to do more of that than I ended up doing, and so this was a bit less zippy than it could have been.

Still, I’m mostly just disappointed by the omission of better potential comparisons, not what’s there. And I can be a bit of a stuffy writer sometimes, so hopefully it’ll be fun to see something less rigid. Video is below, and a lightly edited transcript to follow.

---

So this is my talk, ‘Thought Strewn All Around Us’. And to begin, I want to frame with a quote from Wittgenstein’s philosophical investigations, prop 127:

The work of a philosopher consists in marshaling recollections for a particular purpose.

And that is the frame I want to carry into this talk today. This isn't a sort of a linear argument beginning to end about a thesis. This is about getting together some metaphors, some of which I've used in previous writings, some I haven’t, for a particular purpose. Which is to induce onto you a sort of vibe, and avoiding sort of a rigid and nounified classification scheme about the vibe, and just getting you to see this way.

And the way I want you to see is to think of knowledge as something like a library catalog card, right?¹ That's got the shelf that the book is on, maybe the content of the book, what it's about, but it's not the book. And that the environment all around us is carrying the equivalent of the text of the book forward through time. And so what you think of as your knowledge is more like your index of where to find the knowledge that you're used to substituting for the knowledge so quickly, that you kind of miss that elisioon is there. And so I'm going to use one more quote from Wittgenstein. This is philosophical investigations, prop 78:

Compare knowing and saying:
How many meters high Mont Blanc is;
how the word game is used;
how a clarinet sounds.
Someone who is surprised that one can know something and not be able to say it is perhaps thinking of a case like the first. Certainly not of one like the third.

And that is what we're going to talk about today, that knowing how a clarinet sounds. We're going to start with the world of whistling. Show of hands, who here is able to whistle, who here is comfortable whistling right now for the crowd? Thank you very much. So those of us who know how to whistle — what sort of thing is that, that knowledge of being able to whistle? Maybe on its own, you're starting to think of this in information theoretical terms. Which is to say that some parts of the world are very bland and flat, and it doesn't take many words to describe a large tract of them. Something that's going on in your brain, and we don't need to concern ourselves with the specifics here, is very not flat, is very wobbly, is very specific. And somewhere in that specificity is encoded the “knowledge”, so to speak, of how to whistle.

Now I want you to try whistling again, with your ears plugged. Okay, one of you got better, which is interesting. And he knows who he was. Most of you, though, got worse, and I don't think it would surprise you that you got worse. And so the fact that you can get worse if your ears are plugged, let's think again — what sort of thing is that knowledge of being able to whistle? And now you might think of perceptual information, but still with that magic word information. Which is okay, maybe my knowledge to whistle is more like a computer program. It's got maybe parameters, and then my perceptual module is going out into the environment. And it's pulling in, this sort of information is coming in. And then here's my, you know, listening system, and the perceptual information comes in, and it gets compiled down and the output is a whistle. You may notice that I'm doing weird things with my hands, and I don't have a diagram up on the screen. That's because I don't think this is the right way to think and I don't want to make it easy to think this way. I want it to be hard to think this way. And the reason is that perceptual information, I think calling it information is sort of doing a disservice, because it is sort of — listening is a spot that we feel like we are on more solid ground, that we have a sense of where the variations are.

But an interesting thing happens if you go into Google and you type “why can't I whistle any more?” Okay, if you go and Google, why can't I whistle any more? You get some things are like, you know, have you tried to licking your lips? Have you tried listening carefully, those sorts of things. But what's interesting is what comes right up front, what the people are looking for, which is: have you had dental work done? Have you been to the orthodontist? Because maybe your mouth was in a shape that permitted whistling, you've had some dental work done, now you can't whistle anymore. And so imagine you're one of these people, you liked to go around, you'd liked to go whistle, you went to the dentist. Your mouth isn't the kind of mouth that whistles anymore. Now for the third time, I'm going to ask you — what sort of thing is that knowledge of how to whistle? Now, for you. And you can kind of try to keep patching the ship and say, well, there were the parameters going into the program. But there's also the binding meta parameters about when it is —and now this is like an orphan procedure that the call stack doesn't reach — and da da da,  and so and so.

But I think that a much easier way to think about it is that what whistling is, it's an understanding about a particular sort of environment that you can reach. It's a sort of idea of this environmental space. And that at some point, as your body went forward through time, it happened to produce some lips that were capable of whistling, then some other things have happened, and now it can't. But that your lips are doing some of the thinking about how to whistle. And you can tell that your lips are doing some of the thinking because if your lips change too much, you can't whistle anymore.

And when you start thinking about this way, you realize that there's sort of infinite caveats, right? That you sort of, you can't know everything that you're happening to hold constant for whistling to be able to work, or any other thing. And that's sort of the first key to my argument that I just want to let sit for a minute, is that what you're looking for are things that vary and things that don't, but there's infinite things that could vary. So you're sort of never going to get there by enumerating a list of what can and can't vary. And so what you need, and what life has indeed evolved to produce, is some sort of strategy about how to negotiate this space. And so as an example, if you imagine you can't whistle anymore, and you're really bummed you're walking on the street, and then some little Dickensian street urchin doffs his cap and says “It’s okay, guv, you never forget how to whistle once you learn!” Because Dickensian street urchins can't afford dental care. So the Dickensian street urchin’s knowledge of how to whistle and your knowledge of how to whistle, you can see that it really does happen to be invariant for the Dickensian street urchin in a way it's not invariant for you. And so what we need here is a strategy about how we think about invariances. And when we're thinking about invariances, we're thinking about environments, okay?

So another metaphor for what we might think of knowledge is, is if you have a guitar teacher, and you're trying to learn a song, and they put pieces of tape on the fret, telling you where to leave your fingers, right? That's a guitar that's been tuned, that's a room that the guitar has been tuned in. That's a song that’s sort of, you get close enough, it's good enough, it's recognizable as the song. And so those stickers certainly aren't nothing. They’re a kind of knowledge, but they're a very contextual kind of knowledge. And I want you to start thinking that all knowledge is sort of contextual, on this gradient of invariance. And that negotiating that gradient is just not something we're very good at, we kind of really like the idea of a fact. And because some things ground out in very profound invariances — let's say photons as an example, okay? One photon, you really can just treat the same as any other photon. And we know this because we can tell through evolution that photocells have been around for billions and billions of years. So photocells must just have been working like that. We see a lot of photons coming in, we make predictions about photons, and they're validated all of the time. And so I'm not trying to tell you that sort of everything is unknowable, right? If you kind of just treat a photon as a photon, that works, and if you treat a whistle as a whistle, it really doesn't. And so how do you distinguish between photons and whistles?

And another thing sorry, I, something I missed in my outline before is that another important thing when you're thinking about photocells, is that photocells are reflexive, right? You don't sort of think about am I going to interpret the light in this way or not? It just happens. And so the more invariant something is, the more that we tend to not even think of it as knowledge at all. And so it's worth knowing that with knowing how to whistle, which some people know and some people don't, and is much more contextual, with these conditions — the very fact that we've parceled it out as something that one might know automatically puts it on the more dependent scale of the environment than something like seeing where nobody says “How did you interpret that photon, because I kind of thought it might be red?”

And so not only am I saying that we need to be on the watch for certain kinds of knowledge. I'm saying that it tends to be the more interesting kinds of knowledge that are more susceptible to this problem. Because the more that something is sort of fixed in the environment, the more that you really can treat one x just the same as another x, the more that you tend to do so automatically. And so now I want to kind of switch a bit, and I want to talk about ocean waves. Because with whistling, like I said, the problem is that there's sort of infinite things that could influence it. And this is true. But we're going to simplify the problem of that by looking at just one artifact of the ocean, which is wave size. Okay, the size of ocean waves are very important. Lots of commercial things are at stake. This is a problem that people are thinking really hard about and really need correct answers for. And there's also a thing called rogue waves. And the idea is that it happens to be that waves interfere constructively and destructively. That is, when you've got wave systems going up and down, sometimes the two ups will touch each other and make a really big up. Sometimes two downs go and make a really big down. Right, so you've got a rogue wave, and a rogue hole in the ocean.

And this used to actually be considered a myth for a long time, because we kind of wanted a more regular thing, where somebody says: If I sit and look at the waves, I can see a lot of waves, and I like to ascribe them with an average wave height, and that's that. And then some sailors would say, “Yeah, but sometimes you scurvy dog², you see waves three times the size of the average wave height”, and you’re like, "Go to bed, Grandpa”, you know. And, you know, it's kind of treated as a fish tail until the Draupner oil platform clocked a rogue wave 30 feet up³ and was like, oh, the other waves aren’t 30 feet up. And it turns out that this is just an implication of a chaotic system like ocean waves, that sometimes you can get one big — and in fact, it's sort of a power law distribution where you can sort of imagine, well, what happens if two rogue waves collide? Do you make a roguey roguey wave? And the short answer is that yes, you basically do. But of course, the way that power laws go, the rarity of a rogue wave, you have kind of square that to get the rarity of a roguey roguey wave. That's not what they're called. They're not that frequent. But that's the idea.

Okay. So if you're, let's imagine you're trying to predict the rogue waves, which are of vital importance. Well, what we have now is we have wave forecasts that are accurate to...I saw a few different sources, trying to give an authoritative source on this would disprove the entire ethos of my talk. So I'm not going to. Let's say about 10 days, okay? About 10 days out, you can predict how chaotic the waves are going to be. From how chaotic the waves are going to be, you can get a rough sense of the chance of a rogue wave through some equations. This is a rational enterprise, it's a useful rational enterprise. You don't predict where the rogue waves are, okay? We don't know where one's going to go, that construction and destruction.

And if you wanted to get there, you might think, well, I'll just simulate the whole ocean. But you can see that sort of, the more you care about, not just the average wave height, but the specific location of a given rogue wave, given that waves will knock off of any individual jet, or ship or buoy, then your data set needs to be more and more and more detailed. And eventually, your hard drive needs to be so big and so wet, that it's got to be the ocean. There's nothing else that stores the data with the fidelity you need. And that's why you know, if you're going to Mars, you don't sit and type in the plan and go to Mars, you go set a rough plan and constantly course correct to see how you've gone. If you want to predict the ocean waves, you've got to do it over time, because you don't have the data fidelity.

And now I want to ask you, what do we do today if we want to predict ocean waves 20 days from now? Well, the answer is that you take the dataset of the ocean, you run it in the ocean for 10 days, and that's the output that you feed to your prediction program. And so this relationship between data fidelity and environmental compute, is that the data has these fidelity's with a sort of infinite infinity. As you zoom in more and more, you start needing to know about each and each individual buoy, each individual ship, each individual outcrop to be able to predict the waves. But once you add the dimension of time, then sort of the arbitrary fidelity of the environment becomes arbitrary compute. Because the way that you figure out how all of these things are going to be intermeshed over time, is to wait and let time run them in the environments. And so this…is this is the relationship that I want you to think about the environment as being, right? it's not just sort of, we're gonna get our progressively deeper and deeper magnifying glasses. It's that over time, all of these relationships, the way that their output is stored by the fact that we can just continue to keep looking at it. And so the environment is the compute, which is why I say in a real sense that the environment is the knowledge. Because when we say we understand that we can predict waves for 10 days, that's because 20 days ago, we'd let the ocean run the ocean for 10 days. And so all of these predictions, all of our knowledge, because we are beings in a finite point in time, and we're making predictable artifacts to be used as a point in time, we think about things in a thin slice. But it's only possible because we look at the output of how things have always run arbitrarily far in the past. And so when you're trying to have knowledge, the knowledge you need is how much do I care about what's happened in the past? And you can't give definitive answers about this without looking at the present.And to look at the present is to accept the bountiful gift of the environment, running its own correlations for arbitrary amounts of time.

And so who here has read Nick Bostrom’s Superintelligence? Nick Bostrom’s Superintelligence is sort of arguing about the fear of an arbitrarily powerful AI that is going to produce as an artifact, sort of arbitrarily powerful knowledge in its database. And when Nick Bostrom was trying to think of the limit case of like, what's the longest it could take to make AI? He argues in emulated brains, which is to say that our neurons are information, okay. And you can scan that information into a computer. And if you scanned it at the neuron level, boop, boop, boop, boop, boop, it's a number that's very, very big. But it's not unthinkably big. It's the kind of thing you can't put in a computer today, but it's the kind of thing we might reasonably expect to put into a computer tomorrow. And so Nick Bostrom says, “Well, if nothing else, once we have one of those, we can start doing some computery stuff to it, and iterating and self checking in the way that computers are great at, and that's gonna become super intelligent”.

But here's my question for you all, is that emulated brain going to be able to whistle? [Crowd: Yes, by building the whistle robot that it’s attached to.] Yes. By building the whistle robot that it's attached to, exactly. And so now, will that whistle robot have had orthodontic work done? Well, that's very important, because we know that some lips work, and that some lips don't. And that the act of producing lips that work is an interplay between our genomes, running in our environment, and accepting certain foods and accepting certain orthodontic works, and maybe getting your front tooth knocked out by a baseball and maybe not. And all of that stuff goes into compute whether or not you can whistle. And we don't have to sit and do that computation. We just go, [whistles poorly] “Oh, I got one that works, kind of.” And so we don't have to. And so we have this remarkable information saving strategy we can use, which is just to let the environment do most of the heavy lifting and just find a strategy that works in the environment. The whistle robot’s not going to have that. If the whistle robot’s gonna have to know how to whistle, then it's going to have to know about lips. And you can see that's a harder problem than just scanning everything in your brain that’s attached to your lips. And the question is, how much of what we do is thinking is being buttressed by our bodies and by our environments? And how much of the strategy of emulating a brain is to attach it to simulate_the_natural_world.exe? And how much harder is it going to be to make simulate_the_natural_world.exe then it's going to be to scan the information in our brains and represent it? And I'm here to argue that it's going to be infinitely harder. That whole brain emulation, in terms of making a human being, is sort of 0% of the problem being solved, to just look at the artifact. Because the artifact is situated, and the artifact developed over time. And it's all of that development that you need to do things like whistle.

And to kind of take this and ground this in an actual example that we have today, because we don't have full brain emulation, I want to talk about leafcutter ants. Leafcutter ants are one of the coolest animals in my opinion, because like us, they're also farmers. So an interesting thing about leaf cutter ants is they don't eat leaves, which may seem weird, because why are they cutting them? And the answer is that leafcutter ants are cutting the leaves and marching them down to their hives because they have fungus gardens there. And that's what they feed their larvae on. And that fungus doesn't survive in the wild, right? That is a domesticated fungus that is groomed by leafcutter ants, it's given leaves, it's kept clean, it's kept free of parasites so that the fungus can keep growing and growing. And they feed it to their young.

So let's think about the leaf cutter ants genome. This is another case where the impulse to think of this as just “information”, this knowledge, is really overwhelming because they literally have a genetic code. It’s literally four letters big, four and a half, you can literally look at and say, like ACTG GGC. Okay, this is — we've represented in the way that we haven't yet done a whole brain emulation of human beings. We've certainly sequenced the genome of leaf cutter ants, right? That's happened, that's done, that's fine. But where's the fungus in all this? Because they don't pull the fungus out of their own bodies, right? Fungus is growing on its own. But a leaf cutter ant isn't going to survive without the fungus. And in fact, if a leaf cutter ants fungal colony dies out somehow, they have to go to war with a neighbor. And they go and they raid the fungus chambers of their neighbor, and they take it back to their hive to kick things off again. They have to because otherwise they're going to die out. That's the strategy that they follow. And so something in those twirls of the leaf cutter ants genome, is this idea: check your fungus garden. Put your leaf into your fungus garden. Is there no fungus in your fungus garden? Find another leaf cutter ant and kill them, just kill them and take their fungus for your own, you must take their fungus. Where is ‘their fungus’ in the genome? What is it? Just like we asked ourselves, what is that thing, our knowledge to whistle? What is that leaf cutter ants knowledge to maintain their gardens? Well, it's a pointer. It's a pointer to their garden. And if their garden is dead, it's a pointer to a strategy to find the gardens of their neighbors. And so if you emulated a leaf cutter ants genome, in your computer, you managed to figure out protein folding so much you click click, click click you perfectly build the leafcutter ant. And you click go. What's gonna happen? You're gonna say, “What the hell are these guys eat? They're just all dead.” My little simulation, click, dead dead dead dead dead. That because you need to simulate the fungus, too. You need to simulate the fungus over time. And what happens if you emulate the fungus’s genome? Even if you have the thing that you can't have, which is an environment simulator in general, it's just going to get out competed. You're gonna say, “How is this? It's like this fungus isn't even trying? How is it still alive?” And it's because both of their genomes are pointing to the actions over time of the other versions, right.

So it's all pointers, okay? That's what I want to impress upon you. And that sometimes your pointers are things like photo cells, which are pointing at a photon, which you can reasonably expect to understand the mechanics of, because it acts in a very consistent way over time. And some things are like the leaf cutter ants, which is over a period of millions of years, it's symbiotically evolved with fungus, you can maybe look at two things of fungus and see how much genetic divergence there's been. And if you want to do that work to bring in that knowledge, it hasn't changed too much over millions of years, you can try to reconstitute the past. And if it's something like whistling, then it matters how you were fed, it matters whether you've had work done, whether you've been hitting the teeth of the baseball, it matters on life scale. And some things matter even more rapidly than that. And some things are even slower than that.

But it's all a continuum, right? Anytime you have a knowledge artifact, what your question is, is, what is it situated in? And how can that situation change over time? And so the first thing I kind of want to impress upon you then is this library model, that you have a card index, and you're going to read the knowledge. And because your environments are often very stable… sorry, let me rephrase this. It’s more that your environments are very correlated, right? Something that doesn't happen is that sometimes the oxygen all goes out of the room and comes back. We don't all sometimes fly up and fly down. Lots of things stay very similar over time. And I'm not contesting that. And that point is, when something is relying on those, you don't get orthodontic work done on every single day of your life. You don't get orthodontic work done randomly. So when you say “I know how to whistle”, you can usually whistle tomorrow, and that's good enough. But because that's often good enough, we sort of discount how important this lookup process is into what we call knowledge. Right? We're so good at talking to our environments that we kind of forget we're doing it. And I just want to impress upon you that what you're doing is getting these very, very big library reading lists and reading the books very quickly. But if the books were gone tomorrow, you wouldn't have something. In the same way that something had happened tomorrow and you suddenly can't whistle, and then you're going to Google it. And then you're going to learn this distinction that wasn't relevant to you as a Dickensian street urchin who only recently hit a windfall, but was relevant to other people previously.

So, okay, one more example. Let's say, it's a very, very cold day. And you're wearing a winter coat. Now, from an informational perspective, cold is sort of..things are very similar, right? The world is kind of trying to take your togetherness, your local interestingness, and disperse it. Because it's, you know, that's what cold is thermodynamically speaking, right? It's sort of a lack of information that wants to tear your information apart and spread it around more evenly. And that's less likely to happen if your coat is zipped up than if your coat is open. So when you think about that, your code is sort of meaningful, right? On a universal scale, your coat is meaningful, because your information is able to go into the future more smoothly with the coat. Whereas if it's not, it might get sort of run ragged, frostbite might take some of your skin and say, like, “All this interesting skin stuff is doing, it's just dead tissue now, because I wanted that information for myself. I took it.”

But let's say the zipper on your coat is really fiddly. Okay, it's really hard to do there. It's some weird contraption. And it's kind of half rusty. But it's your coat. So you don't care that nobody else in the world knows how to zip up this coat. Because you just gotta jimmy it, and it's fine. So if you think about this, now if we’re imagining all environmental knowledge being a library, I want you to imagine the back room. Ooh the back room is a mess, stuff is just flying about every which way. In fact, infinite stuff is flying about every which way, because sort of the nature of all infinities being correlated, which is to say, sort of every single trait that something could have, may or may not be linked to another trait that something could have, the back room is much bigger than the front room, and nobody knows how to get there. This is all of the information that just sort of churns, you know, the rock falls, and hits another rock and such in such a way and the rocks happen to know but nobody knows that it's happened.

But your coat, that's equivalent to you going in the backroom and writing a catalog card yourself. And the sort of, the positive side of this meaning — this analogy of mine is kind of down on a lot of eternalistic knowledge as a whole, right? Because my point is sort of all knowledge is conditioned on its terms of invariance. And the fact that this is such a weird kind of wibbly talk, for me, it's because we don't have great languages for, you know, levels of invariance. But the positive side of this is that you make meaning yourself in a very real, and I would argue thermodynamically rigorous way. Which is to say if you can zip up your coat, and nobody else can. And your coat does indeed take this structured information of how you whistle if you happen to whistle and haven't had work done and your favorite foods and how to flip a coin and everything else you know. And since it forward in time more easily, then you have made that coat meaningful from your knowledge of the zipper. And you can see your knowledge, it's an index card, it's not a book. If the zipper breaks, if the zipper rusts too much, if the coat is destroyed, that's not going to mean anything anymore. Just like your knowledge of whistling didn't mean anything once your mouth changed. But it also meant that you, in this moment in time, you are creating the meaning by going to the back room and cataloging something. And saying for these terms, for these hands I have now which have developed through processes I can never hope to fully understand. But I've let the environment run them for me and get me to this moment in time. And in this moment in time, I'm able to zip up my coat. And so please take this out of the backroom and put it on the shelf, because I know where to find it again when I need it. Thank you. 

1. I call these “index cards” in the talk, which led to some confusion. I meant the catalog card that says where the book is on the shelf.

2. This was a failed improvisation; scurvy dog is a term of endearment among pirates ,or at least a put-down to fellow pirates. I wanted “landlubber”.

3. I was way off, it was 84 feet! The 30 feet figure I had in my memory was the historically believed upper bound of how high a wave could get, but the real deal was something else entirely.