From Ants to Active Inference with Daniel Ari Friedman
This is a conversation between Andrea Hiott and Daniel Ari Friedman discussing various themes related to cognition, perception, action, and Active Inference. The conversation starts with Daniel's time with Deborah Gordan studying ants. Then they delve into the understanding of different terminologies and concepts relative to complexity, the individual and collective behavior. They discuss Karl Friston's active inference, relative terms like predictive processing and coding and their application in scientific research. Daniel explains some of the principles of active inference in relation to cognition and perception and how it can be viewed as a scale-free (or better scale-friendly) framework. They also discuss semantics as measurement, and the role of the Active Inference Institute. The conversation is reflective and philosophical, touching on the intersection between cognitive science, neuroscience, and environmental interaction. They come to words like 'service' and 'love' before it ends. Podcast artwork by Daniel Ari Friedman Listen and find the full show notes anywhere you find podcasts.
Daniel's personal site: https://www.danielarifriedman.com/
Research site: https://cognitivesecurity.us/
Postdoc review: • Postdoc in Review (2020-2023)
Active Inference Institute: https://activeinference.org/
Paulo Sayeg: https://philpapers.org/versions/IENPPU
RJ Cordes & Dean Tickles: https://zenodo.org/records/8266281
Lars Sandved-Smith: https://scholar.google.com/citations?...
Relative to way-making research: System 3 representations and a discussion of affordances (a post does not have them) Andrea & Daniel also discuss or mention: scale-free and scale-friendly. Individual and Collective behaviour of ants. Complex systems. Bioinformatics. How we can find real boundaries, and how not-so-real boundaries can be modelled. Uncertainty. Language as a model. The difference between predictive processing, predictive coding, and active inference. Why a fence post does not have affordances. Affordance beyond Gibson. Perception as inbound regularities. Action as outbound regularities. Buckminster Fuller: "Unity is plural and at minimum two." Gregory Bateson: "It takes two to know one." Deborah Gordan Gordon's paper Wittgenstein and Ant Watching Karl Friston Chris Fields Research The. paper: 'Competency in Navigating Arbitrary Spaces as an Invariant for Analyzing Cognition in Diverse Embodiments' Active Inference
Link for Spotify: https://open.spotify.com/show/51TOeA9...
Transcript:
Ants & Active Inference with Daniel Ari Friedman
Andrea Hiott: [00:00:00] Hey everyone, this is just a little introduction to this conversation I had with Daniel Ari friedman of the Active Inference Institute, and I will link to that institute and to some of Daniel's work because he's done all kinds of things. As you'll hear, we start out talking about his work with Deborah Gordon on ants, which is in complex systems.
which is really fascinating, and we lead into a talk about active inference, and along the way, we talk about what is an individual, what is collective behavior, uh, ants are a really good, uh, place to start when you're trying to understand what an individual is, where you can draw the line. Is it the ant, is it the colony?
We don't really mention emergence here, but we kind of dance around this idea. And, um, of course we get into trying to understand what cognition is, which is this, where this active inference model of Karl Friston becomes relevant and the work that Daniel's [00:01:00] doing with the Active Inference Institute.
Active inference says that, what the brain is doing, let's say, is minimizing free energy. And it talks about the relationship between perception, action, and cognition. Daniel lays it out here He talks about how we can imagine, This inbound and outbound statistical dependence. So if you think of something like the brain or the body, you have a Markov blanket, kind of a interface, a boundary, and then you can look at the inbound dependencies as perception and the outbound dependencies as action. The second thing that's really fascinating for me, and that even though Daniel and I agree, I sort of push him a lot on it, is trying to understand a model like active inference as a model, as a map of the territory, not the territory, as a measurement of the process, not the process.
Daniel explains all these first principle ideas of active inference in a way that's very articulate. We talk about the Active Inference Institute that he's developed and [00:02:00] what they're doing there, which is exciting and I encourage you to check it out and look at their at their work and also their papers.
We talk about scale free and scale friendly and all kinds of ideas that can seem, at least for me, it's a it's a bit at the edge of my comprehension at times, but I like that space and I hope you do too. And, uh, thanks for being here. And thanks to Daniel for the great conversation.
Hey, hi, Daniel. Thank you for talking to me about these themes today.
Daniel Ari Friedman: Well, thank you for having this conversation.
Andrea Hiott: Framing it from going into a conversation about active inference and cognition what got you into, to studying ants
Daniel Ari Friedman: I worked with ants from 2014 to 2019 during graduate school. So how did I come to ants?
Well, I was working in a Vinegar fly, also sometimes called a fruit fly, Drosophila, developmental genetics laboratory in undergrad, and was getting also interested in system [00:03:00] science, complexity theory, all these kinds of cool ideas swirling around out there, seeming to connect dots and fill spaces between, give us ways to work with.
Situations of interest that go outside of the theoretical, all these different things I was getting interested in and from the fly genetics space looked what else was out there in insects. They're cool, they're everywhere. Some questions are apparent about insects, whether it's a beetle or an ant or a bee.
Other questions may only reveal themselves through looking for a longer time at a species of insect or at. Perhaps interactions amongst species, insects or otherwise. And so I was recommended to apply to go to graduate school with professor Deborah Gordon at Stanford and ended up following through with that. And Professor Gordon has had a 30 plus [00:04:00] year long term field site in the Arizona, New Mexico, U. S. region. One of the few long term studies of ants, and also I'm sure topics we can go into, but just here I'll say takes a very deep and complex systems perspective on the evolution and ecology and development of collective behavior.
And so pretty much on my first visit out to the field, Walking around thinking, wow, five years of Pogonomyrmex barbatus, red harvester ant. I should come to love Pogo.
And so having the specificity. Of the dissertation and the specific questions that we were pursuing and the entry point of a piece of dirt, an area with a certain set of colonies of one species as an entry point with a long term history opened up avenues that we can explore.
Andrea Hiott: Yes. So, I guess you were mostly focused on the genetics, is that, was that your [00:05:00] main focus when you went in?
Or were you thinking about bigger issues of action, perception, cognition, things like this? Or what was your sort of frame in those early, when you first
Daniel Ari Friedman: started? A great question. So my first email, my reach out email to Debra, Was looking at some work that had connected behavioral variation amongst colonies to their reproductive differences to fitness, which used DNA based methods.
And I emailed asking whether CRISPR. A gene editing technique that was just becoming more used in the fly community at that time had been applied to ants. Short answer was, no, not in 2014. But in 2014, we actually did see that happen in other ant species, not our work. But it just showed how, even in so called non model species, there was an introduction of bioinformatic methods, including like genetics and epigenetics.
We did transcriptomics, [00:06:00] so gene expression analysis on the brain tissue of foraging ants. We looked at neurotransmitter levels, specifically dopamine, and the relationship between neurotransmitter levels and behavioral differences amongst colonies. We did not do any of the kind of epigenetics in terms of histone, Or methylation based studies and did not go too deep into the DNA based genetics, either like a genome wide association study. But to have awareness of the whole stack, even though it's outside the scope of any lab, let alone dissertation, we're never going to get to all the books. Complex systems doesn't mean that we're going to go into every single field and measure every single ant on the ground, but there's something nice to be said and done between having a field site or a field way. And also a broader contextualization for, for a lot of different reasons
Andrea Hiott: so you're already looking for big [00:07:00] patterns. I mean, of course, when you're doing this kind of data analysis and you have to, but you're already focused. On trying to see how that relationship worked. I assume you're looking at individual ants when you did these procedures, but also taking data in a wider sample
Daniel Ari Friedman: individual ants. Isn't that the question? Is the individual ants. taking a measurement of the colony. Yeah. For example, taking a blood flow measurement from the individual, you might abstract over the blood cells and just look at the rate and say, well, that's a single phenotype measurement we're making of this blood vessel.
And then you might also be able to take the red blood cell perspective and talk about it from that side. Actually so much and more rests on individuality, understandings of individuality and individuation. Yes, in our own reflexive self narratives in the human social setting, but then just speaking in the scientific behavioral [00:08:00] ecologist setting, which one is what we're going to call the ant.
So even if we walk around with an understanding that there's multiple ways that we could talk about individuality, whether at the nestmate scale, the colony scale, population scale, and so on, in the way that we speak about it, in the framings that we prepare experimental plans along, and the kinds of assertions or hypotheses that we bring to bear on that specific system of interest, one of the beauties and mysteries that doing something as simple as how do these colonies regulate foraging across and within days, even just with flags and toothpicks and hand counters, there's like a whole lot universe of complexity that the individual and the team can unfold.
Andrea Hiott: I wonder though, like, uh, in terms of where you were setting your framework or your scope, if you already were beginning to sense that it's a little bit subjective, what you call an [00:09:00] individual, or even, for example, some of those questions that you just raised can seem really magical if we start at this so called individual ant level or Some part inside the ant I mean it can just seem impossible to understand or or magical I guess is the word but if you started at some other scale it might not seem so if you took the individual to be the colony or something like that so I wonder where you playing already with those kind of ideas, did you realize that this is a little bit of a subjective thing where we set the scale of individual, group, and so on?
Daniel Ari Friedman: It's a great question. An early tension or question that I remember was thinking, well, when we use keywords or the papers that we're interested, they talk about collective behavior.
No one is saying that behavior is ever really done by something that isn't a collective. We know that bodies are made of cells and so on, or however else we want to describe that relationship with bodies. So what is it that makes collective behavior? And it's like, well, obviously, what's doing [00:10:00] the work is not really collective, but collective is just the shadow of whatever that wouldn't be, would be individual behavior, but people who are studying one rat in the tea maze.
They don't call it individual or collective behavior, they just call it behavior. But also when they study social behavior, for example, even in a laboratory, they call it social behavior, or just behavior. Again, not collective behavior. So, in collective behavior, I saw something like a yearning to potentially not just look at a more multipartite setting, but to explicitly map amongst scales of behavior in its occurrence. So that we could implicitly or explicitly have individual behavior, but then we're focusing on another level of organization. Um, what is collective and what's the individual. And then the way that that space or continuum [00:11:00] gets, um, metal pressed out by the keyword, symbolic, discrete ways that we have.
Which isn't even to criticize them. It's actually to bring articulation and clarity. Then, one other piece I'll add there is a lot of different frameworks. Whether information theory or active inference or other are called scale free. And, which is to say that, that The essence of the framework or the firmware, the kernel doesn't have a system of interest or a scale that it's grounded in.
So you could do a linear regression with X and Y axis being millimeters, or it could be kilometers. So scale free by itself actually does less work than I believe many ascribed to it, because again, a linear regression is scale free. So it's kind of like, well, what else have you done other than just be scale free because numbers are already that [00:12:00] way.
But that's not obviously the end of the account and only recently through. So many different conversations and collaborations. Have I also started to figure out and out the figure with Dean and talk about scale friendly and understanding, well, there's actually an approach or a mediation that is the scale friendliness that allows interoperability and allows us to go from end to end plus one or end to end minus one, not just to be outside of it and make something that's quote scale free, which then always has to be um, scale specified in the last mile to become empirical, but then scale friendly to me was like the connecting piece that took scale free. And filled it in with something that was more continuous.
Andrea Hiott: So maybe you have to tell me a little more about, about scale friendly, but just to pause [00:13:00] for one second.
And I really like what you said about the scale free, because scale free is very useful in terms of opening our minds to the possibility of nestedness and so on. But even when you have something like set theory or category theory or whatever, you are still having parameters. And setting them and um, the linear ness of it, maybe we can like just put this in the atmosphere.
I think we're all trying to figure out ways to get out of this linear way of thinking, which even sort of structures our language and our mathematics and even scale free and all of these terms. And we're all talking about multi scale multi dimensional nestedness, but we still seem to see things linearly.
We can come back to that theme, but but first I want to hear about scale friendly because I that's new to me I think I've I need to read some papers, you have to direct me
Daniel Ari Friedman: as of today, I don't think you'd find a written case of it. Okay. Because it's part of an oral tradition and really in that way, an example of all the kinds of new words and spins and weavings of [00:14:00] words when you say linear, well, that can mean linear, like a ruler or Euclidean geometry or something like straight in line. Although people often don't realize that generalized linear modeling can be both linear and have the appearance of a curve. For example, a lot of neuroimaging is done with general linear modeling, but that doesn't mean that it's being put a ruler over, but another kind of linear, not just the points that are on a manifold is the linearization of speech as a constraint again connecting to the oral traditions. So there's a linearization of symbols on a string or of the production of speech that's a different kind of sequential linearization. Um, and so it's always about reading between the words, unpacking all these different richnesses, like the finger pointing at the moon isn't the moon, the words don't bear the meaning. They're just written on the screen that then is for us to do the [00:15:00] um, enrichments and unpacking of,
Andrea Hiott: um, In that also is time, right? The linear time that plays into all this, to the math, the language, and so on too, but this beginning and end, you know, I think this is also even broader what I mean too, is that we, without thinking, we are looking for beginnings and ends to things in the same way that we're looking for a beginning and an end to what an individual and a community and a group is and all of this.
It doesn't work so well anymore, but it's very hard to think. Without those terms.
Daniel Ari Friedman: Great points. Like what does death refer to? Even if one takes a nuanced concept of individuality, then still, then what is death messaging? Those, um, kind of linguistic notes aside, again, what differentiates scale freeness as something that we construct is like it's an analytical representation. It says nothing. And that's what allows us to take it. Like the paintbrush says nothing. The pen says nothing. So a scale [00:16:00] free framework, linear regression, active inference, renormalization group, pattern languages. So it doesn't even need to be numerical, let alone formal, like category theory.
It doesn't even need to be that, but just scale freeness, maybe even emotions or scale free or some other.
Andrea Hiott: But there's still some trajectory and some parameter in some landscape, some space, whatever you want to call it, that's been set, you know, by language or by the community or the group or something. I mean, the terms, right?
Daniel Ari Friedman: So here are these different disciplines which become embodied and realized through the people who you assigned. into that hypothetical room, the neuroscientist and the psychologist and so on.
And That actually reminds me of one of the philosophical works that Debra wrote, which is about Wittgenstein and ant watching. And so in that, um, one and only philosophy ish paper, [00:17:00] there's a very interesting discussion about how, well, we see something moving on the ground. Let's just, let's just put aside all the questions that lead us to us looking at the ground, but something's moving on the ground.
So that's the individuation in the figure ground question. Okay, so we say foraging is happening. We see a four, we see foraging happening. Okay. That nest mate is a forager. Like that person is bringing something. They are a bringer. Okay. That is to move from essentially a named observation of behavior.
That doesn't mean that it's how the nest mate or any other system thinks about it, but the human observer has called it foraging as an, as a noun. I mean, sorry, as a, as a, well, as a noun construct of foraging, but then that becomes Attached to the corpus that carries it out. So that's why we talk about, well, the nurse and the forager nest mate.
Well, yes, but there's tasks switching and there's a developmental trajectory. It's not [00:18:00] like they're born into one path or another. They all move through this kind of archetypal trajectory, even just in this one little sense. So that again reflects how we go from what we name and observe phenomenally. In our sense making and then through action through speech, but also internal action, which is attention through those kinds of actions. We then reify because it's not like that nest mate is going to say, I'm not a forager in a language that that namer might understand.
Andrea Hiott: Yeah, that's wonderful that you bring that up. It's something I think about a lot, because I feel what I'm trying to do, what a lot of people I feel like we're trying to do is understand, how to be an assessor or a measure from a scale other than your own, which is what we do, right? I mean, we build tools, telescopes, microscopes, all of this stuff to try to do this, um, but it becomes difficult to do.
To then say where the subject and object is and what kind of information we're really finding, for example, like to try to put it back to the ant. [00:19:00] But if you're taking one individual ant and you've somehow, let's just say you've been able to develop a, some kind of device where you can actually be inside the body of the ant and sort of know what it's like to be an ant, get the sensory data of that and somehow, translate it into something that you understand. I mean, it gets connected to how we think of self too, right? Like, how do we start to understand that that Measurement of the ant is within some particular trajectory, um, another position, than the position we've been measuring from. And yet it's only still one possible measurement from that position, right? That gets back to this, how do we get out of this, for me at least, linear way of thinking that it's not that we found the answer of what the ant is, but we have found an answer.
Does that make sense?
Daniel Ari Friedman: Here are a few thoughts on that. I think of two Bucky Fuller quotes. One is unity is plural and at minimum two. [00:20:00] In this setting, that's the observer and the observed, which is also the quantum information setting. And
Andrea Hiott: It
makes me think of Richard Rorty. Sorry, there's no truth, but there's truths.
I think he took that from Buckminster Fuller, but yeah, go ahead.
Daniel Ari Friedman: Uh, possibly, I mean, it's very consistent with just this kind of omni happening, where there's, there's many sensors. Everywhere is a sensor. To itself. And that explains everything from quote physical phenomena like propagation through media, but then also could refer to other kinds of systems that are modeled using synergetics or tensegrity again, scale free frameworks that can be used to say more because they are less opinionated.
And so it's like, the linear regression, active inference, synergetics, complexity, it's not going to have a stance. Can't be expected to, and really wouldn't be desired to be expected to have a position on whether this startup should have this kind of structure or that kind [00:21:00] of managerial structure.
But perhaps what you're alluding to with the kind of disciplinary ways, circling around something is, can we. among other things, shift our attention to the moon, not all the different fingers pointing to it, and then imbued within that, maybe essentially, maybe, um, sufficiently, unity is plural at minimum too. There just isn't a non relational stance. Expressions are relational assertions. They are not stand alone declaratives, except in a very special controlled settings. So just taking a kind of like, fuzziness, crystallizing clarity, and intersubjective alignment, rather than, well, there's actually just this and that, and then we see only blurring of these essences, we can see a sort of what's sometimes called bottom up [00:22:00] construction or arising, but also unfolding and change, of what can be also properly called archetypal or pattern or top down causal influence or whatever it is in a given situation.
At the theory level, we're just building our toolkit. And trying to respect the literature and other ways of knowing and expand the ways that we can talk about sense making, decision making, nested systems, like all of that. Just stamp collecting, insect collecting. Why not have more nuance and expressivity in the language?
Andrea Hiott: Yes, I think that gets to the need for, it's more of a cognitive shift in a way that's happening, you're kind of unsticking from this notion of One answer or, one perspective or so on.
But I also think a lot of what you said can get very, it can, you know, then we get into this trap of reductionism or something like this too, where, or even, um, where [00:23:00] it's just, there's no truth at all. I mean, that Rorty quote was there are truths, but no truth, but it can easily become there's no truths. Because. You can measure anything from anywhere and so anything, anything goes, which actually isn't, doesn't follow logically, but I didn't hear your second, the second quote you said you were going to think of too, but.
Daniel Ari Friedman: Oh, I seem to be a verb. Ah, okay.
The idea that we could even take a process or procedural or dynamical understanding of the self reflexive construct of I, obviously, and also other things. Why not? We don't even experience the visual scene statically. There's a temporal thickness, even to vision moving an object fast. And we know about the blind spot and the vision and the color and the periphery and the isocade, all these other features.
So we just kind of can pick up at the empirical and [00:24:00] swap across and gap across with theoretical and empirical. They're a dialectic because unity is plural and not minimum too. Because we're not going to be able to put it all into one box because collapsing to one box is like a death state. And so having multiple, minimum two, not only two necessarily, but being able to have that anytime we see one, knowing that there's a minimum of two, if nothing else than the agent in the niche, pretty general, pretty good one right there.
Andrea Hiott: One of my favorite quotes from Gregory Bateson, and it takes two to know
Daniel Ari Friedman: one.
There's so many, so many mutations on that. People could take and run with so then it's like well
Andrea Hiott: then what
but we're scientists or I mean I'm more of a philosopher, but I I Was in neuroscience labs and science is supposed to be objective and we're trying to find answers and it does matter so like how do you Take all this that we've been talking about which can start to feel very it starts to feel very, [00:25:00] beautiful and, exciting, but also as if you're getting like disintegrating a bit in, in what's possible.
The lack of being able to kind of draw hard boundaries. I mean, do you ever find that tension or how, if not, how have you, how do you resolve it or how do you deal with that?
Daniel Ari Friedman: First, I'm reminded of Hamlet, melt thaw and resolve itself into a dew and cyclic phase transitions and fractal phase transitions and like that. Oscillation, and yeah, there's a disintegrating, and then there's also a synthesis phase, unless it's a runaway, which eventually must find another oscillation. So, that doesn't mean we can always, at some level or way of the fractal cognitive fractal, we can't always perfectly predict.
where the wave is going to be if it's turbulent or if it's an 100 foot wave. However, we can use real time unfolding variational free energy, recalibration in the moment, and we can also use our anticipation, expected free energy, and [00:26:00] planning as inference. And so we have a lot of the bounds, not to say the only ones or the exclusive ones, but there are some boundaries that can be realized that actually are the shadow or the framing of what might be called a fuzziness.
Like, well, the probability of raining tomorrow might be anything. Well, anything like anything, anything or like a number between zero and one, because we know that that's going to be a number. But just more, um, professionally or pragmatically here, the research project is the container for that uncertainty. And so if somebody can find an arrangement or articulation with like
their research work in terms of how much they're being rewarded and valued and if that is working structurally within the envelope of the project, the work and attention and [00:27:00] dynamics and give and take and all that is beneficial. then that's an adaptive and a lucky and honorable position to find oneself in and help others be in as well.
So at that structural understanding level, that doesn't mean that everything is like fine in the experiment. It doesn't mean everything's fine at home, all these different challenges, but those are the challenges within the bounded. And it's really a state of exception that breaks outside of that boundary.
Andrea Hiott: Yeah. I like it that you started with Hamlet because there's something about, you started with this art, right? Which, or, or this expression that doesn't need to be quantified. In the sense that I was coming at it with this science and, how do we nail things down? And that is kind of already an answer in a way.
And I think where you're talking from is from a certain perspective that I don't know, a lot of people don't have that kind of a perspective, a way [00:28:00] of getting outside of their work enough to see it in the way you just described. I guess I just wonder, did the ants help you get there? Or
Daniel Ari Friedman: again, different ways I could approach this, getting outside of my work, I wouldn't say I did or do, I have made it the work, part of the work, just like scheduling is truly part of the work, and following up, and the before, during, and after of an activity is part of the work. The after and the before, and also from a neurological perspective, the during, like, includes sleep.
That's something, and then, again, this isn't a solution, it's not solving for x, it may even be a precursor step to being able to differentiate letters at all. But those are fundamental cognitive steps, because that challenge is a cognitive challenge, or it just, it can be understood as cognitive today, or the [00:29:00] patterns that are studied in cognitive science related to sensemaking and decision making are some of the patterns, but it doesn't need to be all, that are important for this cognitive slash metacognitive task, which is just one way of knowing and working, it's not any person all the time, and it's many to many just like tasks are in the colony.
Andrea Hiott: So you already saw that as like integrated, the mental and emotional and the scholarly. And so I guess what you're saying is you, you hadn't sort of separated that out.
It was already for you, part of the enterprise from the beginning, which is, that's, that's interesting.
Daniel Ari Friedman: I, I don't know. There's, there's that evaluation of what did happen for me, which Again, who knows? And it can be, it can both and and then there's the transferability and, and that one thing that it makes me think of, though, from what actually [00:30:00] did happen was like being an undergraduate doing fly genetics.
it was clear the time and the space and the way that things had to be done. Oh, here's 10 trays of tubes that need to be tipped. It's just 144 times 10. It takes this long. Here's how long to schedule it. Listen to podcasts, listen to music, talk to a friend. Um, and then in another kind of level or way, graduate school.
Is like, okay, well now there's some more agency, especially up front, like situating the qualifying exam or finding a question that's motivating or that's really curious. It seems like there's going to be epistemic value. And then that goes from there. So like that is the work itself. And some work is very structured.
And there's definitely [00:31:00] complexity about, like, showing up and being, like, authentically how we are. And also, there can be authentically repetitive work that is essential.
Andrea Hiott: I like that. It's almost like the training was meditative and That can be, yeah, that's, that's a really fascinating way to think about school and the work and the lab work and stuff, almost as a meditative practice that then can scaffold some of the other things.
And especially when we get into these huge, seemingly chaotic or whatever, this, these bigger parameters of where something can seem fractal and you can just like, as you were saying, there's just so much information or there's nothing that like one person can't deal with all of it. Um, if you already have that grounding in a sort of meditative structure, I can see how that would really help.
Daniel Ari Friedman: Let me say a little more on that. So yeah, the actual pipetting and the fly flipping and ant watching and whatever it is that every [00:32:00] other person in their playings and hobbies and trainings experiences.
Um, but then it made me think about like, in the disciplinary way, there's kind of like the elevator of just like more or less hard, or it's kind of like a tree, but whether, you know, it has some fundaments and then it kind of bifurcates into the more, um, arcane research sub sub fields. And so education is kind of spanning that whole tree, different disciplines.
There's a whole forest forest for the trees and it's like, whether it's it's a undergraduate level class or research presentation or a language model generating a performance art piece about a scientist, whatever it is, if it's the right thing for you to be in that listening environment.
Then, with attention, it can always just be absorbed with awareness. [00:33:00] If not used as a catalyst, it doesn't even mean that what they say is accurate, that you would have said it that way, that they said it the best way that they could have, that it was most recent, that they mentioned this recent paper.
Did they, you know, do the list ABC or CBA? That's all critique. And there's a time and a place. And there's a way and maybe there's even a job, but like for actually, yeah, but for being there and just like, I just want to enjoy this lecture and just, I'm not going to go into this area.
I'm not going to be tested. Those modes, but also even designing our own modes of attention and figuring out paths to get there. Again, those don't really rely, that, that doesn't really rely on any specific research finding.
Andrea Hiott: No, but it does, you do have to know you can do that.
I think that's kind of gets back to what I was trying to say about some of it's just realizing you can do this, what you've described, like that actually within itself is quite a big thing. And some of this actually just is an example of what [00:34:00] we were talking about at the beginning, that there are different perspectives.
You and I, I think, have entered the world and experienced the world in really different ways. And I can't say what that is, but I feel like if I'm going to generalize really big for me, it's more that I, the world was this kind of very emotional, artistic, I don't know, overwhelmingly emotional thing that I then sort of, through kind of hard work, realized I could structure or something. But I guess there's just different ways of, of waking up to, the fact that you are a cognitive subject and then that you can actually start to take action about how you arrange and scaffold, all that input it can be overwhelming, to want to experience the world in complete full and always, if you haven't, uh, learned a way of structuring it, not necessarily by being in the lab and learning that, but, um, however you learn it. But, uh, I want to kind of turn towards. Yeah. [00:35:00] These terms of cognition and sense making and active inference. So you've said since making quite a few times, we've been talking about active inference and so on.
We've been talking about cognition and this, this does relate to all we're talking about already, but with the ants or any subject matter, the action is very, pretty clear, right? From our perspective, as we're measuring and assessing it, perception and cognition, these words start to get a little more confusing, right?
And, um, as we're going to lead into a talk about active inference or even predictive processing or, or whatever, we're talking about action, we're talking about perception, we're talking about cognition, I wonder where when you started to think of if you started to think of those as different, or if you've thought of those as in a way that we've been talking about.
And as I think of it as kind of different approaches to understanding the same process rather than necessarily like different processes, perception, action, cognition, maybe you can just tell me a little bit about what your memory is of how you started distinguishing between those or not distinguishing between those.
Daniel Ari Friedman: [00:36:00] yeah, a lot of, a lot of levels and stages
when I'm today saying the things that were said earlier or heading into this more explicit section on active inference, I really do think about the active inference ontology as a dialect, in this case, a dialect of English and, but it doesn't have to be English. It could be a programming language. It could be another natural language.
So that helps make the toolbox. For me, separate from the deployment of the tool. Like, I may or may not know what every single hammer and screwdriver on the wall does. Maybe some are just painted on, they're just jokes. You know, maybe some are useful for multiple things. Maybe there's uses for one that haven't ever been done before.
But certainly every bike coming into the shop is a new bike to fix. Every new question is a new one to address. basic or advanced, totally knowable, open and shut, [00:37:00] seemingly only one answer or two answers, totally open, just an evocative question. So how did, especially looking back, now understanding active inference, only in the limited way I do, how does that help me understand how I was looking at perception, cognition, action, and impact in the niche?
Well,
the first thing that's not active inference that comes to mind is wondering why we had all of these ant colony simulations from myrmecologists from insect and ant researchers making simulations of different ant settings. And then also a lot of interest in the computer science and in the pure math space of like ant colony swarm optimization.
So there's a lot of interesting models proposed for decades. However, seemingly, I don't know to what extent this is empirically borne out, but we could find out, it didn't feel like people were ever following [00:38:00] long term on one type of model. Like a group would tend to introduce different versions of a kind of research theme, but with like a first principles architecture each time, maybe by new students, um, or new technology coming into play.
But I wasn't seeing too much versioning, scaffolding, and application direction Or synthesis, like somebody saying, well, actually these two simulations can be understood as basically composed of, um, you know, special cases of, of this more general way of thinking about it through the general framings of agents that cybernetics brought into play, it was very clear.
For things that are behavioral, we can think about them in terms of their inbound and outbound statistical dependencies. If something doesn't have inbound and outbound statistical dependencies, it's like It's either not there for you to [00:39:00] see and or not having a feedback effect. So it's not even just that it could be discounted.
It's like it will be discounted because it won't be within your perception
especially in the last years with the new tools that have been developed in active inference and the free energy principle. There are examples now of first principles descriptions of perception, cognition, action, impact, and the niche. This isn't to say it's the only first principles, or that first principles would be preferable, but this is a total, synthetic, first principles unifying approach to those diverse phenomena.
And specifically, it involves a four fold partition where there's a blanket, Or interface or a boundary. And there's two types of states on the interface that just coarsely have dependencies going in different directions. And so with respect to a given system of interest, we can [00:40:00] just analytically associate the inbound dependencies with perception, outbound dependencies as action.
And then everything that happens on the other side of the blanket as cognition. So that's not an account of perception, cognition, or action, or impact in the niche for any specific system. But that particular partition, the fourfold partition, then is like a tool that we can take to now use that dialect and formalism on the GPU.
In the computer, or a bacteria in the glucose gradient, or ourself at this time, or this team at that time. So, it's just like, the tool that we're bringing to the table, not saying that the tool is the answer, identifies the problem. All of these different features, once the tool is clear, it becomes a lot clearer at a meta level.[00:41:00]
What is the tool? What's the problem? What's the solution? What's the issue here? Not even saying that there are other tools, but I just, I think it's a really fascinating area. And I really do, um, think that there are some differences in the stance that active inference and potentially other related, uh, approaches may already be offering us.
Andrea Hiott: Yeah, I agree. I like it that you started out talking about these words as language, and that language itself is a system that we're using to assess. Let's say, whatever, a process, that's going on and that none of those would be really like stuck to the process necessarily so much as the way that we're assessing it and we have agreed upon in, in inputs and outputs and regularities, be it mathematical or be it words.
So are you sort of saying that the, the actual math or maths. That Friston or, or others have come up with, um, relative [00:42:00] to predictive processing, for example, are you saying you could almost see that as a language to that's approximating how to assess and measure a larger process in a way that we can all communicate about and agree upon in a sort of cybernetic, uh, out of that lineage connection, communication, finding common language and so forth.
Daniel Ari Friedman: That's quite the question or hypothesis, certainly as a technical instrument. It has some of those features that might be qualitatively summarized from earlier in our discussion, like unity is plural at minimum too. So having relationality and the observer observed and dynamics on both sides of the screen, that's a really cool feature.
It's scale free and also potentially scale friendly. So meaning not just that active inference doesn't presuppose centimeter or kilometer scale free, scale friendly in that we can make an active inference model of an estimate. [00:43:00] And then 10 years later, we could say, okay, now let's do multi agent simulation.
So we just scale friendlied it into collective behavioral setting. And then someone says, well, I think we should have a tree. Okay, cool. Now ant colony and tree. So that is composable and interoperable. So applies category theory very well in that area. And this is, uh a way that we can talk about those settings.
There may be other ways to talk about and work with those settings. These may not be all of the settings. Just being able to talk about something is not equivalent to making an adequate expression about it, especially in the embodied setting. So there's still a million clear caveats. Are there unknown unknowns caveats?
Yeah, probably. But every, these are fundamental, but that's just like when people are like, well, maybe numbers don't represent things. It's like, okay, but that's not really a ding against any type of linear regression or active inference. That's a very, you know, that may be a [00:44:00] valid point. Maybe your valid point might be a useful point.
Rug pulling all of numbers isn't doing a knock against any given model. They're just different scales of analysis. And so. There's more to say there, but as our communication bandwidth, not just the digital bandwidth, but being able to speak remotely is quite helpful, but as our bandwidth and alignments and our reference frames through language and education and culturation, all these things increase, especially with sentient artifacts and augmented synthetic intelligence.
That's unprecedentable.
Andrea Hiott: Yes. And then we're really layering the themes we're talking about, but to continue doing that, I think, yes, it's all true what you said, but when it comes to something like the field of neuroscience or the field of cognitive science, or when we're just trying to understand what is cognition, do you not think that someone that these what you just described is very uh democratic and fluid and what I described in a way is [00:45:00] just one way of possibly understanding a process uh do you think we actually use it in re in the discipline like that or is it more that there's only going to be one answer this is what cognition is because when, when you're talking, I agree, but I also think it doesn't, that's not really the way people talk about it in the literature.
It's more, we're trying to find an answer with good reason because we want to solve some issues that people face relative to mental health, for example, or brain, diseases or whatever. So, um, yeah, I, I just wonder like, where do we, where do we, uh, pull in and out philosophically
Daniel Ari Friedman: Yeah, a lot of good points. Well, especially as viewed from the inside and like the bottom up, ascribing intentionality to the scientific literature or field, either it's not a relevant level of description, or at the very least there's many degrees of freedom on whether the field of neuroscience wants [00:46:00] anything.
But this also returns to individuality and what things can want things. I find that it's very challenging and nuanced to separate out, like, again, the finger from the moon. For example, in the active inference area and in the Oxford English Dictionary, so it's not like this is a rare take, a term like sentience might be used to describe adaptive anticipatory or allostatic intelligent action, but not necessarily phenomenal awareness.
But sentience can be used in both ways. Or belief has a narrower technical meaning in Bayesian statistics, just like attention, surprise, and all these other features do. But then there's this, we need a minimum of two, because belief, attention, precision, were brought by analogy [00:47:00] from cognitive science into statistics.
Just the variable, the precision on a variable, or the attention on a variable, isn't the human experience of attention. And so then it's like, but why do we have the same symbol? It's like, what's wrong with symbols being min two? But we need to have a little bit more nuance sometimes so that it's not like semantics, but everything is semantics, even if it's narrative.
So But people often mean that it's like a word game saying that it's just semantics. It's like throwing up your hands like, Oh, it's just a card game. It's like, yeah, we're playing a card game. It is semantic. Now let's be serious and sincere and fun and inclusive and all of this, but let's do it then.
Let's just do a subscript. If these are the two definitions or let's have. an open source repository. And if you want to use consciousness 14 or 19 or whatever, [00:48:00] just make it clear. And then we'll know. However, these have, first off, these technologies don't always exist, or are not always in use, and they may fundamentally, but certainly in practice, have very high cognitive overload and burden.
And stakes. Frankly, yeah, and high stakes.
Andrea Hiott: It's not really just a game. This, what we, how we define cognition and how it trickles down to becoming, I mean, it has everyday consequences on all levels, whether why we fight each other, why we don't fight each other or how we think of our.
The boundaries of self or how we address mental issues, things like this. So gosh, I mean, maybe I can try to get into a lot of the stuff that I was thinking of when you were talking by asking this, when you're talking about active inference, um, how strict are you to something like predictive codings, relationship to active [00:49:00] inference and mental models and things like this?
Or are you, like, when you're using that word at the Institute, um, because you have a lot of different people come and speak from different ways of defining that word, but I guess, like, how do you see that? Are you strictly tied to something like Helmholtz Friston predictive coding tradition or what?
Daniel Ari Friedman: Great questions. So I'll give a more limited answer than a broader answer. So the limited answer will be about the relationship between active inference predictive coding and predictive processing. And then the second is I think a more general and important question too, about like how we select platform and what a field or whatever, what have, what X Institute is.
Um, so to the first point, active inference is a unified approach to modeling perception, cognition, action, and the impact on the environment. The specific way That sophisticated cognitive entities are modeled is with [00:50:00] an internal world model or generative model. that includes anticipation or predictions about the environment.
So that is often called predictive processing, whether it's a single layer or a multi level predictive processing architecture. On the quote, lower levels, you have more sensory information coming in. And then on progressively like higher levels, you have abstracted representations. So computationally, if not biologically, so predictive processing is referring to the bottom up flow of sensory inputs As surprises, but the top down imposition of priors now predictive coding takes that last piece, which is encoded as surprises and says, what if not just the architecture being laid out that way, but what if the messages.
that were being conveyed were about the difference between the, um, expectation and the [00:51:00] observation. So you could have a predictive processing system where each level sends down, here's what my prediction is. And then that, then this one sends, well, here's what my observation was. That wouldn't be predictive coding, though, because in predictive coding, only the delta is being sent up, only the difference between them.
So predictive coding is kind of like a wiring choice. Predictive processing is referring to any kind of predictive model with incoming sensory information. Very common cognitive architecture and active inference is a very unified approach to model those kinds of systems, but also it's more expressive.
And so it is in the tradition and the lineage of predictive processing and predictive coding. However, it can be used to express. other forms of cognitive models. The broader question, though, is very interesting in importance. And, uh, I hope we can elucidate and be clearer on it. But the few things that come to mind [00:52:00] first, there are fractal containers.
So if somebody has an organizational psychology idea that's like cool, and it relates to like keywords that are in the active inference ontology supplement. So it's like they're adjacent. This person is not talking about Bayesian statistics. However, they're talking about something that is like part of our broader topics of interest.
Then creating like an org stream series and just saying, yeah, let's have that conversation and make that two way street act like it's an active inference. Enthusiasts first time hearing about your type of organization. And for those who are familiar with your kind of organization, let's have the conversation.
It's not a test that brings that in. Let's have multiple people hear multiple voices. So being able to actually have the genre and the format and the, um, location of deposition under our agency as well means that there [00:53:00] can be spaces that are more informal. more adjacent, less proximal to active inference that absolutely can be held and have been some of the most important.
And we also have like, Oh, this paper just advanced to capacity in active inference. So then however it is that it did that, however, it is now entering in to the, the mesh of the developing framework, like in the last several years. Since we've begun the Institute, topics like Bayesian mechanics and category theory applied to active inference were not even on the radar, but now they're fundamental components in our education.
And so basically being able to have the continuum and then having great community and ecosystem. Who first
Andrea Hiott: used the term active inference? It's a great
Daniel Ari Friedman: question. Like people have used them. I've done some limited keyword searching. [00:54:00] There are people using the words in conjunction in the early 1900s, but not in the way that it's being used today.
Using active inference to mean a process theory that describes cognitive activity compatible or consistent with the free energy principle is happening since the early 2000s. We're going to have many, many fun literature analysis times ahead with speech and text analysis taken to a new extent. And so I very much look forward to a much less anecdotal and immersive linguistic understanding of different languages.
And different syntax. Yeah. With overlapping semantics, obviously. Yeah.
Andrea Hiott: And, and doing that something I think about a lot is the original Helmholtz, when we're talking about free energy principle and all that we, we are
talking now more [00:55:00] about this literature of the 2000s that's Friston oriented but it goes back to they call it unconscious inference but when you really look at the german it doesn't even have to be inference unbewusster schluss and That's more like a conclusion or something, not an inference, which gets to a point that I wanted to talk to you about, which bothers me a little bit, is the idea of inference sounding already kind of meta to me.
It sounds, and especially maybe this is because I studied predictive processing, predictive coding and so forth from in neuroscience. So it's what comes to mind for me are these images of brains with the priors up here and the, you know, sensory input coming in from the body here. And like this whole, it's all brain oriented and it's all representational and like taking those representations to be inside brains, which frisson doesn't necessarily do. even though he models it that way. But so it's, it's already ambiguous, but when I just think of it, it's these models of brains and it's models [00:56:00] in brains. And it's inference in the sense that to me feels like, Awareness of your knowledge or knowledge of knowledge, which bothers me because I think of something like the ants, who are also, I think, in the same kind of loops and even that recent book, with I think some of your colleagues, um, talking about what is active inference, it's generally a thought of as a process that doesn't start with humans, of course, in the human brain architecture, but as, you know, this ongoing kind of process that maybe has resulted in that, but that's not like the top tier or something.
So this, I guess I'm just throwing this out to see if you have some way of elucidating it because it bothers me a lot. The idea of inference sounding so top down and so already contemplated versus trying to think of something like an ant just making its way in the world, and having this, these, these loops or more, I would say spirals, not really like circular action perception, um, ongoing ways of just processing their world sensorially that we can then assess as cognitive.[00:57:00]
So that's a lot, but
Daniel Ari Friedman: Okay. I have a bunch of kind of short, funny responses. Only in the spirit of comedy and inspiration, of course. Yeah, that's good. So first, how any English word hits any person on a given moment should be like a tiny, tiny feature of professional argumentation. Of course, as a friend, I think it's always interesting to hear somebody's unpacking because it conveys richness that goes way beyond the topic.
But when it comes down to what a term should be, One person's taken in one moment is like it's like a self ad hominem. It's like a self own second inference is what the algorithm does. It is a Bayesian inference algorithm if you use a Bayesian statistical engine. So map not territory. This is cartography.
We're saying that the algorithm.
Andrea Hiott: Yeah, I have to stop you there though because that's already the mathematical like we were saying That's an assessment of something.
Daniel Ari Friedman: Yeah, but the active inference, when we make it in, [00:58:00] um, RX and for software package, it is doing inference. That's already
Andrea Hiott: representational, right?
Doesn't mean the ant is inferring, which is perhaps a line.
Daniel Ari Friedman: It doesn't mean any, any system of interest. It is, it's probably wrong to say it's doing active inference in the way that the program is implementing it. Now there are analogies and that is what makes maps useful for territories, but the framework itself narrowly certainly does inference.
But whether a bacteria in the gradient or whether a metronome on the table or whether a neuron or brain or a glia or anything like that do inference is, is it literally a separate question than what the field should be called?
Andrea Hiott: But connected. And I love, I have to say, sorry, I know I'm interrupting you, but the map territory thing is great that you said that because that's exactly what I mean is as we assess, you know, it goes back again, we're layering from the beginning.
As we assess another position, of course we can say that's inference, but that doesn't mean it's inference from that position, I guess is what I'm trying to say. And once we use these mathematical [00:59:00] models or computer models, of course it's inference because we've set that structure. Um, but go
Daniel Ari Friedman: ahead. Yeah, yeah, certainly.
It's inference from the constructed model that humans did create. Um, yeah, people certainly highlight the mammalian brain a lot, not just because it's an important system or region, but like. Any time I see, for example, chapter five in the 2022 textbook, it's like, well, let's give the examples. Well, all the examples are from the mammalian neuroscience specifically.
Andrea Hiott: Doesn't it start though saying this doesn't come with, start with human brain architecture or
Daniel Ari Friedman: What's so interesting in the way you say that. First off. It did start with human brain architecture. Humans created active inference. The second thing is the particular partition certainly does not rest on any specific system.
This is again, the key point, which is active inference says nothing about any system of interest. It's any, any more than a [01:00:00] pen says anything about which language or which word you're going to spell. And so active inference is, but, but that bridge is often welded or paper mached over, or just averted. Because maybe it's, it's another understanding to think that theories say things about systems of interest. I'm sharing it how I see it, not just as an expression of how I see it. To me, it's very clear that the theory of linear regressions being developed and people developing decades of ways to do linear regressions, including on dynamical and chaotic systems and nonlinear systems and all that. Those decades of development said nothing about any given relationship between a behavior and a disease. And so, but it's, but there's the regression right there on the paper and it's easy to point at. And so similarly being able to separate active inference for what it is, which is a scale free, which is a system independent [01:01:00] framework and method and tradition from anything that we want to say about any organ.
They're two separate things, but they're complementary in praxis. And that's why often it's more clarifying to do like one behavioral experiment and analysis. Like, how tall do I estimate the people are? The next three people, I'm just going to guess their height and then I'm going to fit a Gaussian. Okay.
That's what I did. That can show more about the scientific process than spending all this time going into theory. Or learning history, but then abstracting or disconnecting or course graining from that fundamental experimental based uncertainty.
Andrea Hiott: Yes, and I guess to kind of riff on that even a little more, what maybe bothers me is that we sometimes we, I don't know, it seems to me from my perspective that we sometimes take the model as if it's the [01:02:00] biology.
Um, I don't know, like, I, I feel like we confused the representation, the language, for example, or the math or the model with the thing that we're assessing in the same way that we might assume inference and in the way that we were talking about. And I, I think that matters. And it's part of trying to.
shift the framework in the way that you kind of brought up at the beginning of understanding it as a framework that can be turned into many, in many dimensions and scales. I don't know. It just, there's something about, for example, thinking of the model, having to be in the brain or the brain itself, even having to be necessary to understand this process that matters for me, which is.
Because I, I think, aren't you, are, are you also trying to understand how this continuum plays out, uh, in all kind of forms and scales of life and species rather than just having a [01:03:00] model that applies, only in one area? I mean, isn't that part of the Active Inference Institute's kind of notion of translation too, or even your own work?
Daniel Ari Friedman: Yes, and not that it's the only one, but we have one of those. Which is the scale free framework of active inference, which says nothing about any given system, and then every last mile for us to flourish into, and live and die, or ride or not, about anything. Which is what some people sometimes also highlight about free energy principle as a theory of everything, with a space there, because It's not giving an account of like, what some other theories of everything project themselves or assert themselves to do, which is to provide a totalistic account of all.
But rather, we can provide our account of [01:04:00] any thing. And if something cannot be identified as a thing by that modeler, it won't be included in the model. And that's often called a deflationary approach, because It is more procedural, it's more descriptive of a scientific method, or an analysis method, rather than something that explicitly talks about cosmology, or human psychology, or anything like that.
And that's why it's a first principles account. Because it brings us from 0 to 1, or 0 to 2, to be able to utilize, not because it gives us some grander scaffold to hang anything.
Andrea Hiott: Yes, I agree. And of course that a lot of the criticism of active inference, which I'm sure you've heard, is that it's way too general and so on. Because it's like, okay, yeah, [01:05:00] there's something very exciting about the words active inference even. It's just there's something very exciting about the notion of it, um, especially because we are able to become aware of our own inference.
And so there's these kind of layers, but I still, I feel like I want to push it a little more because there is a reason that it gets into all this math and to predictive processing and predictive coding. And, uh, I mean, If we take seriously what the idea of cognition and trying to understand it and, and, and trying to understand that it's something like an insect and something like a plant and something like a human all have cognition without necessarily attributing the same cognition to those different systems, which is usually what happens, right?
We say, if we say they all have cognition, we, we are assumed to be saying it's a cognition. The ant and the human are thinking about themselves in some way. And I find that is very, something that I think could be [01:06:00] clarified and improved. And to do it, you really need some kind of specifics, so, I guess, what I want to ask you is, how do you think of cognition? When you were studying those ants, did you think of them as cognitive? Or do you think of cognition as, in an almost more dualistic sense of, uh, it's not behavior. It really is, um, something other than behavior that we might call only like mind.
Uh, not action itself and behavior.
Daniel Ari Friedman: Great questions. There might be some systems where those differentiated accounts are indistinguishable and there might be other systems that those are highly distinguishable accounts. Some work over the recent years by Lars Sanvid Smith and others. framed attention as internal covert action. To me, that was a big jump because it showed that [01:07:00] instrumentally we could analytically study action on both sides of the blanket without wondering whether it was experienced.
Now, our experience may be a fundamental mystery. It may not be. It's something I've also thought about. Also, I wouldn't want my personal experience or any one thing's experience to be the beginning and the end of an explicitly intersubjective framework. So when I hear cognition, I see it as a keyword, as a token that is very coarse grained, always needs to be described, whether it's cognition in the abstract system independent, or if it's system specific, if it's abstract, then the sport And the joy is making it system independent, just like linear [01:08:00] aggressions, don't mention anything specific about any disease.
But then if we're going to be deploying cognition about a given thing, usually the claims are specific enough where it can just be operationalized. And it's a non issue, at least not a theoretical issue like foraging in the abstract versus one meter away from the nest entrance that moves something that is.
in theory, open to in practice, actual. And I just find that a huge amount of settings where cognition is used to point at, whether it's cognitive science or cognitive security, in the actual description, there's a lot more clarity and it's a different kind of thing than the abstract, which is Essentially always open, always could be generalized, or always could have a counterfactual.
That is its own kind of [01:09:00] conceptual clarity. Being able to propose new variants of a theory is what we want. Generality is what we want. Expressivity there. And then there's a different dual or complementary imperative. If we were talking about the cognition of any given thing, if someone says, Well, that fence post It's like, it's doing cognition.
It has no affordances, no capacity for action over the timescale. I'm modeling it and it's doing a perfect inference on where it should be. And it wants the same thing as I want. And someone's like, well, cool. I just think of it as the fence post being there. Okay, cool. So then cognition did nothing in that setting.
And also, if we look deeper, we find that there's many settings where behavior cognition and these very coarse terms are just being used for indexing and evocation not being, at least I don't understand them or interpret them when I hear them [01:10:00] as being assertions about the theory more generally, but sometimes even within a sentence, people are using things in multiple different ways.
Studying different creatures in, in the world, and then also, going to this level of, Active Inference Institute and humans speaking about all of these things.
Andrea Hiott: Have you seen some kind of continuity, um, in patterns? I wonder what you think of something like just taking any agent in the way we expressed at the beginning, and from that agent's perspective, as it makes its way through the world, however it's doing that, in terms of sensing the world and making something of it, or That, I think, is continuous with making sense, too.
I actually think we can start to understand that as a continuity. Our experience of the world actually being continuous with something like the ant sensing the world and making its way through the world. How does that feel to you when I say that? Or what do you, what comes to
Daniel Ari Friedman: mind?
Few things. The broader [01:11:00] context, even though I saw many different arrangements of pixels on the screen and many different positions of nest mates on the ground, gravity was in effect. The US dollar was in effect. The logistics of scheduling were in effect, transportation, bodily movement, sleep. So overwhelmingly, there was continuity. In approach, given the buffering or the containing a variability that was hand in hand with developing that capacity and developing the capacity to describe it, which, of course, separate things, some people look at the blanket, the interface, Markov Blanket, however they want to say it, and they think it separates, when actually it's what intermediates, and that is the dialectic of the mediator.
It both intervenes between and it connects, and there's no reason to come down to one side of separate or connect, [01:12:00] in principle because it can't be done, and then in practice because it's irrelevant because they're describing the same
Andrea Hiott: element. Right, and, and would you agree also you could redraw the Markov blanket, so to speak, from a different, if you, if you change to another agental or agent based perspective, wouldn't the Markov blanket itself also change without, uh, without erasing anything that you found from that position as truth?
Daniel Ari Friedman: Absolutely. There is a, a blanket around every node. There's an interface around every element. And which one we tag as internal, which then dictates which ones are inbound, which ones are outbound, which ones are external, that is a secondary choice that's not on the table. Um, patterns across the areas, most recently working with RJ and others, where we've explicitly taken the pattern language approach to composing our understanding of cognitive settings that help make it [01:13:00] exoteric.
What various people just in this one corner were grasping at implicitly, and that's happening in many different ways, which would only be right and relevant. And so, yes, there are some similarities between being out on the physical field and being in an intellectual field. And, Another piece of the puzzle that, like so many pieces, is only understandable later is Mike Levin and Chris Fields work on competence as navigation in arbitrary state spaces.
It's like, so how am I going to walk from Colony 623 to 897 without disrupting anything? That's, and then also, how am I going to get from this paper to this paper? How will I get from speaking to this person to that person? Or how will I get to knowing what I want after [01:14:00] this happens? Like those can be modeled with the same hand and the same pen, same computer, same analytical framework.
Andrea Hiott: Just to push a little bit more What about in, in, in the subject matter when you're studying the ant do you, do you see any patterns in literally the way that the ant Or the ant colony makes its way, which I think of as cognitive, and the way us here in these online team spaces, or you've read, written about war, these kind of spaces that seem mental or emotional or somehow distinctly human, uh, the way we make our way through those.
Have you seen any continuity and patterns in those what seem like really different spaces? I mean, that speaks to Levin 2
Daniel Ari Friedman: it should be and it will be unpacked for so much longer. But I think about nest mates using the rate and type of interactions to make local decisions.
[01:15:00] People getting local interactions about which stocks to be bullish on or what kinds of events are happening with what frequency where so rate and type of local interactions and local agency that then gets deployed to whatever ends. It is. That's one. Another important pattern that has complete continuity is between the stigmurgic.
The niche modification and the semiotic niche modification, whether through chemical pheromones, semi chemicals, or whether digital niche modifications in our information environment, that there's this, um, articulation between what is the pheromone on the ground. Or the digital data on the hard drive, and then the sentient entity responds and unpacks that pheromone distribution responds and unpacks that information distribution.
And so being able to separate where are we talking about the dead information? [01:16:00] In the niche, and then with that being crystal clear, because it usually is pretty clear, now what is the complexity of the body and the mind? But let's start with what we can already say is the modified niche, and that opens the door to a much more open investigation of what that active entity's cognition is, which doesn't even need to be stated.
And it may be helpful to go in with a minimal assumptions of what that active component is.
Andrea Hiott: That makes me think of something like art or Shakespeare or digital space in terms of the new world of social media and looking at that also from that same angle that you just applied in terms of like pheromones or trails or trajectories that are followed more and more or broken or even with the foraging idea of what's close by and [01:17:00] What's far away?
I have to ask you one little clarification question, and again, it could be a whole episode by itself, but you speak of affordance a lot, and a lot of people do in this kind of world that You know, some of the, the worlds we share, but there's rarely any talk of Gibson.
And when it comes to perception and action, I mean, um, that's, you know, of course that word comes from an important history. And I just like, do you, do you see, do you think about Gibson? And do you see that as co like making sense with, um, active inference and these bigger themes that we've already talked about, or do you see that as just.
You know, you've taken that word and it's in a different context.
Daniel Ari Friedman: Great question. Here, I will point to the work of one of my intern colleagues, Paolo Saeg. And what Paolo did was in the English and Portuguese corpus, investigate the term affordance. And I believe the conclusion is clear enough for me to work with, which is that while [01:18:00] Gibson and the ecological psychologists highlighted the experienced Potentiality of relational action and in their usage of affordance, what we can call affordance to within a schema that was later made Gibson at all highlighted the experienced perceptual component of affordance, whereas inactive inference.
Affordance is used more along the lines of policy from control theory, which is to say capacity for action, still relational, but it deemphasizes the perceptual. And so actually the problem that the ecological psychologists were setting up to do was why certain capacities were perceived. Whereas active inference removes, let alone the consciousness and awareness, it doesn't even need to be Perceptual at all and so it picks up with the word because the relational capacity for action is the key [01:19:00] phenomena it left
the baggage of being used to account for perception, and then just begging the question, well, why does the one that come to attention end up being the one that's selected? And so, it's absolutely there in the citation trees, and it's in some of the papers too, and on a last note there, I advocate for a move away from last names, and the ideas, and the specifics of the literature citations.
will be much clearer than invoking keystone names. Who will still get more than their recognition through the bibliographic and the meta analytic, but there's nothing about a Markov blanket that's helpful to somebody who doesn't know it. There's nothing about Bayesian statistics or any of these kinds of name drops.
That catalyzes education and accessibility and [01:20:00] rigor, except in a very small and cultured set of people.
Andrea Hiott: I think they are name drops. I mean, there are ways of us showing each other what we've read, which isn't always like a bad thing, even if it's annoying and an annoying thing, but, you know, we do need names too.
Daniel Ari Friedman: Partitioning set, partitioning set on a graphical model. Instead of a Markov blanket or or some other formally equivalent expression, it's
Andrea Hiott: actually much more clarifying. If you say it like that, you can actually know what it is just by hearing the term.
All right. Well, I, want to go back to, the active inference Institute, but also your postdoc work, a word that comes up a lot in there is service. You know, or serve. I think it's like serve, act, infer. Isn't that your, the motto of one of the, of the active inferences to do?
And also in your postdoc, okay. Sorry. And then in your postdoc work, uh, service was a really big. I think if I'm remembering correctly and you do a lot of engaging team with teams [01:21:00] and you, and you look at this a lot. So I, I, I guess I'm wondering where that comes from and why that's so important to you and how that, how you see that fitting with these other things we've talked about.
Big question.
Daniel Ari Friedman: I. I try to do slash be the zeroth level response to that curiosity and don't necessarily know what can or should or must be said beyond that, but I'm very happy and honored and blessed to be able to be where and how I am. And I respect other people's journeys and works in ways. So let. all the action and inference and service flow from there.
Andrea Hiott: Is it a way of of opening, making sure things remain more open ended somehow? It's very important. It's a stance almost. That isn't something that you hear in this world a lot. So I'm sorry to push you on it, [01:22:00] but
Daniel Ari Friedman: it's, well, the open endedness with action inference, we already have too much open endedness. So if anything, this actually connects it to, to what that's one piece. Another aspect is the two stroke engine. Okay. So the framework is unity, but also at minimum two. So we have the two stroke engine across the boundary. So now we have the information engine, but is it going to be a dud? Is it going to blow up in our face or is it going to be a real something that we can use to do meaningful work from information entropy and, and perspective divergence, but if it is going to be an information engine or a synthetic intelligence risk engine or whatever aspect it is, then where are we going to go, who's going to decide, how will we know having something that does is a capacity, but then for and [01:23:00] how and by what must be, even before realizing any given stage of whatever, as a capacity, the whys continue to implore a deeper view on the technical why.
Andrea Hiott: Very good. It opens up to the, the word that we don't dare speak, but the idea of love, um, not something often said in science, the openness and the connection, but also the,
the beyond what can be said with words to just go to Wittgenstein since the name did come up to be silent about that, which we cannot speak or whatever.
Daniel Ari Friedman: It can, it can be the case that love is a map of a territory and we can bring all of that to bear. And then we really. Need to be in the field and it's like any amount of theorizing. Oh, love is a [01:24:00] cognitive pattern or as a historical object. It is all of those things. And it's a pointer and it's a referent and semiosis. It's like, yeah. And that whole sphere is like one mode or regime or culture or way. And there aren't even just two ways. So then it's just an opening. And yeah, we could probably have an opening for dignity, for honor, for any word, truly any word, especially if you're linguistically oriented, why not some of the most evocative and meaningful spaces that are like vast and relevant enough to just be like, you can just pour into it and flow through it forever.
Andrea Hiott: Do you find passion in what you do? Are you passionate for it? Is it Have you ever had to cultivate that kind of a motivation or [01:25:00] I guess I'm, I think this state, the active inference and service is some, there's a balance there too, in a way of staying motivated and in work.
Um, so yeah, just last thoughts, if any of that comes, comes back around to how, how there might be some kind of multi scale, uh, fractal nature of, of thinking about, about the passions that. Stay with us through our lives and our work.
Daniel Ari Friedman: It's really a great question. I'm not sure how much I can say or even internally Describe, but I think it's a huge Question and it's a team and a civilizational scale project and it's a personal and a family and a professional Project and so through that kind of plurality and be like they're just aren't shut doors Like even a data measurement is just [01:26:00] the opening.
And so convergence and divergence, and then how, how do we way finds in that? And focus on like the dialectic of the light and the dark with our flashlight or like our foundedness and our unfoundedness and all of that, which will always be there, whether we're in the brightest room or the most totally mapped space or the real unknown, unknown state of exception, all those features.
So that's part of, I guess, our culture and way. And why it's so fun to explore it with you. And I'm sure to hear from all the others that you'll talk with, I would love, you're, you're curious about how we invited people to the Institute. I want to connect the dots amongst love and philosophy minimum of two and amongst the guests. And that's [01:27:00] the real. social and the enacted, that's not just somebody prompting a language model to make something fluent, but hollow. So I wish you well on this
Andrea Hiott: journey. Thank you very much. I wish you well on your journey too. And thanks for, for talking about all this and opening new spaces for people, I'm sure.
Daniel Ari Friedman: We did it now when it was now. We said it how we said it then.
Andrea Hiott: Wonderful. Thank you, Daniel. I hope you have a great evening.
Daniel Ari Friedman: Okay. Peace. Bye. Bye.