Latest / No Way Out / AI’s D&C Cycles Accelerate: Harmonizing Agents and The Big 'O’rientation with Mahault Albarracin, PhD
Transcript
- 0:20Brian "Ponch" Rivera: Hey Moose, today is September 3rd 2025.
- 0:23Brian "Ponch" Rivera: You and I had a conversation earlier today.
- 0:26Brian "Ponch" Rivera: We'll talk about that here in a second.
- 0:27Brian "Ponch" Rivera: I started today rereading this book here called the Spatial Web by Gabriel Rene and Dan Mapes Fantastic book and one of the things that popped up in there was information warfare and I would start scratching my head like I didn't read that the first time I read through that, but that's one of the things I looked at today.
- 0:45Brian "Ponch" Rivera: You and I had a conversation about the hill, so non-speaking autistics.
- 0:50Brian "Ponch" Rivera: We have some connections you personally, me in the periphery amazing connection there about outside information, collecting intel from there, and I came across a couple of things that I want to share on the screen.
- 1:00Brian "Ponch" Rivera: One is from Robin Carhart-Harris.
- 1:03Brian "Ponch" Rivera: What he wrote today is he wrote really that's not what I want to share.
- 1:06Brian "Ponch" Rivera: Anyway, I'll just read it out to you.
- 1:08Brian "Ponch" Rivera: He wrote that in the future, even more than now, what we call spiritual and mystical, will one still be called spiritual or mystical, and two also be called physics.
- 1:18Brian "Ponch" Rivera: All right, that was pretty amazing there.
- 1:20Brian "Ponch" Rivera: And then I came across something in the last 24 hours and that is hackers getting into Jaguar Land Rover production.
- 1:30Brian "Ponch" Rivera: That is of very much interest to a lot of our folks that are listening to our podcasts and what we do on cybersecurity.
- 1:40Brian "Ponch" Rivera: And of course there's a New York Times article today that just came out in the last four or five the publishing of Destruction and Creation.
- 1:49Brian "Ponch" Rivera: September 3rd 1970 is the publish date, so fitting.
- 2:16Mark McGrath: Wow, Only you would know that Moose it's also my grandma's.
- 2:23Brian "Ponch" Rivera: it would be her 113th birthday.
- 2:24Brian "Ponch" Rivera: So there you go.
- 2:24Brian "Ponch" Rivera: So, without further ado, I want to introduce our guest today.
- 2:26Brian "Ponch" Rivera: Her name is Mal Abarassan.
- 2:27Brian "Ponch" Rivera: She is up in Quebec, is that correct, mal?
- 2:30Mark McGrath: That's right.
- 2:30Brian "Ponch" Rivera: And I have a question for you.
- 2:31Brian "Ponch" Rivera: We built up some context.
- 2:33Brian "Ponch" Rivera: Can you help us unpack that a little bit by giving us your background on how you became interested in artificial intelligence?
- 2:41Mahault Albarracin, PhD: Sure, Coming from the social sciences.
- 2:43Mahault Albarracin, PhD: There's a bit of a subtle thing here.
- 2:47Mahault Albarracin, PhD: We're always told what we do doesn't matter as much, unless there's math, it's not going to work, it's not useful.
- 2:52Mahault Albarracin, PhD: So that's the underlying motivation.
- 2:55Mahault Albarracin, PhD: But perhaps more objectively, I have worked with a lot of paradigms in social sciences.
- 3:02Mahault Albarracin, PhD: I studied sociology, I studied anthropology, biology, I studied a lot of fields in my specific curriculum, and I found that there were recurring things, things that spoke to a form of mechanism, but the mechanism was never really outlined properly.
- 3:20Mahault Albarracin, PhD: It was always quite narrative, always descriptive.
- 3:23Mahault Albarracin, PhD: So post hoc, you look at something, you analyze it and you say this might be the way that it happened.
- 3:29Mahault Albarracin, PhD: But you can't make any predictions off of that.
- 3:31Mahault Albarracin, PhD: And so it left me a little bereft.
- 3:34Mahault Albarracin, PhD: And so I started looking into the ways that we could sciences to this notion, because as soon as you come up with a mechanism, what you're coming up with is a sort of static way to categorize things and therefore exclude certain perspectives etc.
- 3:54Mahault Albarracin, PhD: And so, on purpose, social sciences try to do away with this exclusionary framework, and there's a reason for that.
- 4:03Mahault Albarracin, PhD: But I started reading into ecological science, varela, I started reading about predictive processing, andy Clark, and I found that their approach to how we connect to the environment, how cognition connects to culture and manifests in material reality.
- 4:25Mahault Albarracin, PhD: Through reading things like Karen Barrett and a neo-materialism, you find that there is a way to account for diversity, account for the different ways that reality can diffract, whilst also being predictive and precise in your language, and maybe connect across paradigms as well.
- 4:47Mahault Albarracin, PhD: And so then I started reading into active inference and I met Pierre Poirier, who became one of my PhD advisors, and he said you really should talk to Maxo Ramstad and Carl Friston because you seem to be touching on something that they're also trying to develop.
- 5:06Mahault Albarracin, PhD: And so that's how I got more and more involved in more generally programming, machine learning and active inference more broadly, and through the world of simulations we were able to do just that to do predictions in the realm of cultural social sciences and an account for the possibility for that diversity, and therefore not just put people into a clear box, but more like work with semi-prototypical and edge ideas through distribution.
- 5:35Brian "Ponch" Rivera: Well, that sounds like a familiar story, moose, from John Boyd, looking at different sciences and bringing together disparate ideas.
- 5:43Brian "Ponch" Rivera: Mao, there's something I want to ask you Today in the New York Times article or guest essay, I believe, there's something mentioned in there that you're familiar with, with your company, and I understand you're the director of research.
- 5:56Brian "Ponch" Rivera: Is that correct?
- 5:58Mahault Albarracin, PhD: I'm the director of research strategy, so I work alongside the other director of research.
- 6:03Brian "Ponch" Rivera: Okay, so in that article and I'm assuming you read it have you read that article today?
- 6:08Mahault Albarracin, PhD: I've read it.
- 6:09Brian "Ponch" Rivera: Yeah, can you talk a little bit about that?
- 6:11Mahault Albarracin, PhD: Absolutely.
- 6:12Mahault Albarracin, PhD: I mean, I know Gary Marcus.
- 6:13Mahault Albarracin, PhD: I agree with his vision.
- 6:14Mahault Albarracin, PhD: Ultimately, I think he's putting the finger on the correct problem, which is that scaling forever is not the right approach.
- 6:23Brian "Ponch" Rivera: Let's talk about that.
- 6:24Brian "Ponch" Rivera: I heard this analogy years ago.
- 6:26Brian "Ponch" Rivera: I've read Jeffrey West's book Scale.
- 6:29Brian "Ponch" Rivera: You know that's pretty interesting on how things scale in nature.
- 6:32Brian "Ponch" Rivera: For Moose and I, we believe you cannot scale a linear process like a linear OODA loop.
- 6:38Brian "Ponch" Rivera: I think Elon Musk mentioned at some point that in order to go to space, you can't scale combustible engines.
- 6:44Brian "Ponch" Rivera: You need to do something different.
- 6:46Brian "Ponch" Rivera: You have to think a little bit outside the box.
- 6:47Brian "Ponch" Rivera: So can you talk about what is wrong or not necessarily wrong?
- 6:53Brian "Ponch" Rivera: What is the disposition of today's AI landscape?
- 6:57Brian "Ponch" Rivera: What do they have right, what do they have wrong, and what does the future look like with regards to scale?
- 7:02Mahault Albarracin, PhD: So what they had right was that a lot of what we care about has to do with attention.
- 7:09Mahault Albarracin, PhD: You don't want to constantly have your model, take only one perspective.
- 7:16Mahault Albarracin, PhD: You want your perspectives to be dependent on a given context and therefore meaning to shift according to that text and therefore meaning to shift according to that.
- 7:27Mahault Albarracin, PhD: Not everything you say is meaningful in the same way, given, say, the hierarchy of topics that you're really embedded in at a given moment.
- 7:31Mahault Albarracin, PhD: What they are getting wrong is two things.
- 7:35Mahault Albarracin, PhD: One, they don't have the right notion of embedding.
- 7:39Mahault Albarracin, PhD: Embedding needs to be about an embodiment within a moment and a spatial, temporal reality.
- 7:47Mahault Albarracin, PhD: You need to know, or think you know, something about the thing you're talking about, have a true belief, or at least some degree of a truth statement about a belief, and then you need to figure out why that statement is true relative to the rest of the world.
- 8:03Mahault Albarracin, PhD: What caused that statement?
- 8:05Mahault Albarracin, PhD: What does it relate to?
- 8:07Mahault Albarracin, PhD: So this gives you a structure, and that's not present currently, or at least not explicitly present, in the transformer architectures.
- 8:14Mahault Albarracin, PhD: Another thing is that and that's something that Carl says all the time a correct model is probably a very simple model If you have to make it bigger and bigger and bigger and bigger, just to account for potentially a bit of contextual reality.
- 8:31Mahault Albarracin, PhD: You're probably accounting for too much.
- 8:33Mahault Albarracin, PhD: You're making too many connections.
- 8:34Mahault Albarracin, PhD: You're just trying to cram as much as you can within one given model, such that you think you're accounting for everything.
- 8:42Mahault Albarracin, PhD: You're not understanding anything.
- 8:43Mahault Albarracin, PhD: You're just making sure that you have some data points that you can pull from at any given moment, and I think that's fundamentally wrong and that's what Gary Marcus was also pointing towards.
- 8:55Mahault Albarracin, PhD: We have the ability to understand reality by having some core, say, structural priors about reality.
- 9:03Mahault Albarracin, PhD: Our brains are made this way, most animals are made this way.
- 9:05Mahault Albarracin, PhD: Just our very embodiments speak to that structural constraint that allow us to perceive things in a way that is useful to us.
- 9:14Mahault Albarracin, PhD: You don't need to perceive every pixel all the time at once.
- 9:17Mahault Albarracin, PhD: You really only need to perceive the things that are relevant to your survival, or at least to the scale of reality at which you're going to need to be cognizant, such that you can take actions that are useful for your life.
- 9:29Mahault Albarracin, PhD: Anything beyond that.
- 9:31Mahault Albarracin, PhD: You're probably taking in too much information and we can speak to different kinds of pathologies that are like that right, people who take in too much information all the time right.
- 9:40Mahault Albarracin, PhD: They're hyper aware or aware of the wrong elements because they didn't know what to pick from.
- 9:45Mahault Albarracin, PhD: Does that answer your question?
- 9:51Brian "Ponch" Rivera: Now I'm smiling.
- 9:52Brian "Ponch" Rivera: I want to check in with Moose here in a second.
- 9:53Brian "Ponch" Rivera: The conversation we had about non-speaking autistic children lines up very well with what you described, you know, on the show we've talked about.
- 9:57Brian "Ponch" Rivera: When you read an article, you don't read every single letter and every word.
- 10:12Mark McGrath: You kind of predict it, you kind of alluded to that, that.
- 10:13Mark McGrath: But, moose, I want to hear from you, man, uda r, or let me throw the thoughts.
- 10:15Mark McGrath: Well, I mean, I think the time is now to publish that article finally and, uh, decimate that.
- 10:17Mark McGrath: But what I hear mal describing is ooda loop sketch.
- 10:19Mark McGrath: I mean, I feel like john boyd, basically, would align with everything that she just said.
- 10:24Mark McGrath: Um, and as I'm sitting there listening, I mean I can just see it, I can see the sketch laid out and everything.
- 10:30Mark McGrath: So, no, I mean, do you think Ponch?
- 10:33Mark McGrath: I don't know.
- 10:34Mark McGrath: No, it all resonated with me.
- 10:36Mark McGrath: I thought she was visually painting.
- 10:38Mark McGrath: I thought, mal, you were visually painting the OODA loop sketch.
- 10:42Brian "Ponch" Rivera: I imagine she's thinking about the free energy principle and how that sketch, the basic sketches out there, like the perception action loop and the way you see it in a lot of books and articles right now.
- 10:51Brian "Ponch" Rivera: The way we look at it is we just extend it through the Boyd's OODA loop where we show prediction, we show planning, we show the perception, we show what Stephen Kotler calls pathfinding, which is really intuition, habits of mind, that autonomic response and or finger spritzing of fuel, that intuitive feel.
- 11:11Brian "Ponch" Rivera: Moose and I have been kicking around the idea that one of the most under reported, underutilized part of the OODA loop is outside information, and that was that connection back to non-speaking autistic children that are connected to the hill or whatever they may be connected to.
- 11:22Brian "Ponch" Rivera: But my guess is, when folks have some type of experience and flow, they create something.
- 12:28Brian "Ponch" Rivera: They create something like the spatial web.
- 12:29Brian "Ponch" Rivera: Maybe Carl Friston sees something with his work in fMRI and then creates something new based off of being present.
- 12:36Brian "Ponch" Rivera: So, mao, can you build on that some more and then help us unpack what artificial intelligence may look like here in the new feud?
- 12:45Mahault Albarracin, PhD: I mean I think the word is in the grapevine right, Like I think everybody's pointing towards agents, and what they're really pointing towards when they use the word agent is trust.
- 12:53Mahault Albarracin, PhD: I know it sounds silly, but an agent is something that has an ability to act on the world in a predictable way and, ideally, in an accurate way, and by that I mean you can predict something will go wrong all the time.
- 13:05Mahault Albarracin, PhD: That's not useful Right now.
- 13:07Mahault Albarracin, PhD: I can predict that if I ask ChatGPT to calculate something for me, it won't be able to do it.
- 13:13Mahault Albarracin, PhD: The other day I was trying to use one of the services from a payment system that was supposed to allow me to create quickly SQL queries and I swear to God, I used it for two hours.
- 13:26Mahault Albarracin, PhD: Within two hours, their own in-house service that was supposed to be fine-tuned could not give me the right results and I didn't have that much data.
- 13:34Mahault Albarracin, PhD: I had 87 data points.
- 13:35Mahault Albarracin, PhD: They should have been able to give me something right.
- 13:38Mahault Albarracin, PhD: But that's the thing.
- 13:39Mahault Albarracin, PhD: Because they're not really anchored in any sort of stateful belief space.
- 13:44Mahault Albarracin, PhD: They don't have a sense of what causes what they don't know.
- 13:47Mahault Albarracin, PhD: So we're going towards real agents that are either embodied within edge devices or robots or simply that have an embodiment in the sense of a Markov blanket, which means they have a degree of understanding of a domain and the relationships of that domain to the rest of the domains, for example, and these agents will have an ability to navigate causal graphs.
- 14:13Mahault Albarracin, PhD: I think this is where everyone is going Benji is going in that direction, Lacune is going in that direction, Carl Friston is going in that direction.
- 14:20Mahault Albarracin, PhD: This also allows you to potentially make predictions about how safe a given agent is going to be, or a given clustering of agents, ie different kinds of agents that collaborate together.
- 14:32Mahault Albarracin, PhD: Because, yes, you can make a prediction about, say, one single policy space, but what happens when you compose them?
- 14:39Mahault Albarracin, PhD: What are the emergent effects of multiple agents that you couldn't necessarily predict by only looking at them individually?
- 14:46Mahault Albarracin, PhD: So these are the kinds of things that we're going to have to research there, I think in our future.
- 14:51Mahault Albarracin, PhD: But once we have that, you have the ability to make everything smart.
- 14:55Mahault Albarracin, PhD: You have the ability to have real digital twins that can actually think through their connections.
- 15:02Mahault Albarracin, PhD: You have the ability to make large-scale predictions to understand an ecosystem and its sustainability.
- 15:07Mahault Albarracin, PhD: So once we have all that, we also will have the ability to really leverage the data that we produce.
- 15:13Mahault Albarracin, PhD: Right now, most of our data is junk.
- 15:16Mahault Albarracin, PhD: Not that it's not necessarily good although it isn't but even if it was, we can't do anything with it because it's decontextualized.
- 15:24Mahault Albarracin, PhD: It doesn't speak to anything.
- 15:26Mark McGrath: Ponch, the other thing I just jumped in my head backwards planning sure looks silly, doesn't it?
- 15:31Brian "Ponch" Rivera: Oh yeah, backwards planning is something that some military members live by in all domains.
- 15:36Brian "Ponch" Rivera: So in complex environments they try to start with the future in mind and work backwards, and the problem with that is, in complex systems you can't find a relationship between cause and effect until retrospect or till after something happens.
- 15:50Brian "Ponch" Rivera: Now I want to ask a question about agent, agency and interactions, if you're ready for this.
- 15:53Brian "Ponch" Rivera: So we understand that in complex adaptive systems, it's the quality of interactions more so than it is the quality of agents.
- 16:01Brian "Ponch" Rivera: And, with that being said, when we look at humans and, potentially, machines and robots working together, is a human considered an agent and does a robot have agency.
- 16:12Brian "Ponch" Rivera: Okay yeah.
- 16:13Mahault Albarracin, PhD: So there's an interesting point actually here, which is, ultimately, you never have access to the internal states of any system.
- 16:20Mahault Albarracin, PhD: You can have access to, maybe a representation of it, a way that you think it might be working.
- 16:26Mahault Albarracin, PhD: You might have a very high confidence over what you think is going on, but the truth is you don't.
- 16:32Mahault Albarracin, PhD: So if you think of an agent that way that has a mark of blanket internal states and external states, then everything that has the ability to take actions is an agent.
- 16:43Mahault Albarracin, PhD: Humans will be agents under that robots, ais, animals and the beauty of that is that that forces us to consider how robots and ais and other kinds of systems might have to do theory of mind over us and over each other, because if you don't have access to the true inner workings of anything, the best you can do is sort of map over them and hope for the best that that map gives you accurate predictions, and inevitably there will be a time where it won't, because there's, you know, volatility, variability, because there might be a bug, there might be an issue that you couldn't plan.
- 17:23Mahault Albarracin, PhD: There's always going to be something that will slightly shift away from what you expected, and so you need to be able to consider that whatever model you have of something will need to be updated at some point and your predictions, your theory of mind, predictions, eventually will have to be refined.
- 17:40Mahault Albarracin, PhD: So every human will be an agent in that scenario.
- 17:43Brian "Ponch" Rivera: And we are agents.
- 17:44Brian "Ponch" Rivera: Now, One thing that I was told years ago as well, and we still believe this to be true, is that we shouldn't apply engineering approaches to human systems.
- 17:50Brian "Ponch" Rivera: What I understand you're doing is you're applying natural intelligence or how living systems actually, or the way we understand them to function applying that to AI.
- 17:59Brian "Ponch" Rivera: Is that correct?
- 18:00Brian "Ponch" Rivera: Is that what you're trying to do?
- 18:02Mahault Albarracin, PhD: So Carl's active inference and frenergy principle is based on first principles.
- 18:07Mahault Albarracin, PhD: And so if you think of first principles, going from math to physics, to biology, ultimately, if you want to base your artificial intelligence on the only real example that we understand of intelligence and even then that's not very well defined, right Like there's still a lot of arguments and a controversy around the term itself, nobody's really got the definition Then you have to consider the ecological definition of intelligence under the human format, and then you can sort of say well, you know what it's in continuum, animals have it, some plants have it.
- 18:46Mahault Albarracin, PhD: You can even argue that some systems have it that aren't necessarily considered by most people to be intelligent systems.
- 18:53Mahault Albarracin, PhD: Mike Levin has some really interesting ideas about that.
- 18:57Mahault Albarracin, PhD: Generally he thinks intelligence is in the relationships between things and I tend to agree with him and that means that you can extend consciousness and intelligence to things that don't necessarily register right now as intelligent or conscious.
- 19:10Brian "Ponch" Rivera: Can we, can we?
- 19:11Brian "Ponch" Rivera: I want to pause on that, moose.
- 19:13Brian "Ponch" Rivera: How's that connect to McLuhan?
- 19:16Mark McGrath: Well, I mean, all technology is an extension of some human human faculty, right, human human faculty, right?
- 19:30Mark McGrath: So like, if ai if we're looking at ai correctly, it should be an extension of our, our cognition, I would think, right, I mean, like I always, we've talked about before of ai being an extension or a way to augment my orientation, but it still doesn't negate my own human orientation.
- 19:41Mark McGrath: But the medium becomes the message.
- 19:43Mark McGrath: The electronic media is, the is overworking our central nervous system, but in this case it seems like it does become an extension of us, right, like that's.
- 19:54Mark McGrath: That's kind of like if you saw those notes that I put on substack where people you know they're constantly shitting on ai for whatever reason.
- 20:00Mark McGrath: It's cheating, it's plagiarism, it's's super Google, whatever it is.
- 20:03Mark McGrath: But when you go back and you read things like Buckminster Fuller's education automation and Isaac Asimov talking about it, it's actually supposed to expand our creativity.
- 20:13Mark McGrath: If done right and done correctly, we should be expanding our creativity, which ties back to Boyd, which would be that it's augmenting our orientation, which ties back to McLuhan, which means that our sense-making capability is being expanded by this technology.
- 20:27Brian "Ponch" Rivera: Here's a thought.
- 20:27Brian "Ponch" Rivera: If that's true, large language models would make us dumber if we followed them, because we're following an engineering approach, a circular OODA loop, if you will and I don't necessarily want to say dumber, but it won't expand our global consciousness but something that follows an ecological approach, a natural intelligence approach, could help us become more human, and I think I brought that up with Dr Bray.
- 20:48Mark McGrath: I think that's what they were talking about, like what Fuller?
- 20:51Mark McGrath: Yeah, and I think that that's what Boyd I think was the prototype or the proto of this.
- 20:57Mark McGrath: I think that that's what he was getting at with the learning capability when he mapped out OODA loop sketch that these technologies as they come and they become more and more pervasive, that, if done right, that's what it's doing.
- 21:09Mark McGrath: In other words, it shouldn't be replacing humans.
- 21:11Mark McGrath: It probably replaces you know median jobs or you know meaningless things, but like it should be augmenting our creativity, it should be augmenting our engagement, it should be augmenting our understanding of complexity.
- 21:24Brian "Ponch" Rivera: Yeah, so we should become better humans as a result.
- 21:27Brian "Ponch" Rivera: Mal.
- 21:28Mahault Albarracin, PhD: Even or something larger.
- 21:29Mahault Albarracin, PhD: So the way I like to think of it is imagine we are cells.
- 21:33Mahault Albarracin, PhD: Your cells in your body do not compute all the information of all your body.
- 21:37Mahault Albarracin, PhD: They really compute some local area and they understand what they have to do relative to your local area, which is why, if you place a stem cell in any given place, it'll start becoming anything, and in fact, there's some studies that show that you can even sort of revert a cell and then make it become something new in a different spot.
- 21:54Mahault Albarracin, PhD: So that's super interesting, right.
- 21:56Mahault Albarracin, PhD: So think of us that way.
- 21:57Mahault Albarracin, PhD: You don't want to lose your own independence and your autonomy, or at least the degree of autonomy that you want, but there's only so much autonomy that you want.
- 22:09Mahault Albarracin, PhD: You don't want to be telling everyone all the time what to do, right?
- 22:14Mahault Albarracin, PhD: So so we as a species, anyway, are bounded by a degree of computation that we can apply to the world anyway.
- 22:24Mahault Albarracin, PhD: So think about, like, okay, assuming I'm this now, I don't want to compute everything, but I do want to be connected and I do want to gather some insights from perhaps some higher level system, like I do want to understand how to better navigate the world, the planet, other people that I don't have a direct connection with, but I also want to maintain my attention to things that are super relevant to me.
- 22:47Mahault Albarracin, PhD: So then, maybe AI allows us to develop this sort of super consciousness layer that right now is a bit I want to say reactive, right, because there isn't, I don't think, any sort of layer here that takes action right.
- 23:05Mahault Albarracin, PhD: Culture is reflective, but not quite proactive.
- 23:09Mahault Albarracin, PhD: It's more like it jostles us back and forth, like we are in the sea that directly feeds from us right, like we are the ones that ultimately were the cells, and our survival is represented by this overarching system.
- 23:32Mahault Albarracin, PhD: So I think it will allow us to not just become better humans, but perhaps even a new kind of system wherein we coexist better and become better coordinated towards some greater, larger, more interesting goal that on our own we couldn't really fathom.
- 23:50Mark McGrath: Yeah, you're getting it to.
- 23:51Mark McGrath: That sounds like it.
- 23:52Mark McGrath: Sounds like that what you just said.
- 23:54Mark McGrath: That sounds like Teilhard de Chardin.
- 23:55Mark McGrath: That's the noosphere, right?
- 23:57Mark McGrath: Is that?
- 23:58Mark McGrath: Is that what you're?
- 23:58Mark McGrath: Is that what you're?
- 24:00Mark McGrath: Yeah, which heavily influenced Boyd, and there's a lot of connectivity with McLuhan as well.
- 24:08Brian "Ponch" Rivera: And there's also the adjacent possible, the idea of emergence, complex adaptive system.
- 24:12Brian "Ponch" Rivera: The whole is greater than the sum of the parts.
- 24:14Brian "Ponch" Rivera: All these things are connected and what I've seen over the last few years is the neuroscientific community is driving more towards understanding complex adaptive systems.
- 24:21Brian "Ponch" Rivera: We've seen that with Stephen Kotler's work with Carl Friston, we started with complex adaptive systems and now we're moving more towards understanding complex adaptive systems.
- 24:23Brian "Ponch" Rivera: We've seen that with Stephen Kotler's work with Carl Friston, we started with complex adaptive systems and now we're moving more towards neuroscience to understand that from a Boydian perspective.
- 24:31Brian "Ponch" Rivera: At the end of the day it comes back down to the quality of interactions and this actually kind of leads us to another topic, which is teaming or teamwork, human-agent teamwork we talked about that before.
- 24:41Brian "Ponch" Rivera: But agent-agent teaming, or human-human teaming or just teaming, means that the whole is greater than the sum of the parts.
- 24:48Brian "Ponch" Rivera: It's the quality of the interactions, it's the protocols that we put in place that allow us to become greater, again, greater than the sum of the parts.
- 24:57Brian "Ponch" Rivera: So the protocol, the way I understand it, for active inference agents is this web 3.0 or spatial web.
- 25:06Brian "Ponch" Rivera: Can you talk a little bit about that, mal, and how that may bring us closer to this global consciousness?
- 25:12Mahault Albarracin, PhD: Well, it's not to say that's where it's going.
- 25:15Brian "Ponch" Rivera: Okay.
- 25:16Mahault Albarracin, PhD: I think that's what's possible and I think that if it were to make us something greater, that could be it right, I'm not saying that that's the objective, that's my interest, but I wouldn't put words in Same.
- 25:28Mahault Albarracin, PhD: But in any case, it will allow us to have a more interesting kind of connectivity.
- 25:38Mahault Albarracin, PhD: Right now, we are hyperconnected in a way that we don't control.
- 25:43Mahault Albarracin, PhD: That is not the want, and to connect to collect different kinds of data, none of them are connected.
- 25:57Mahault Albarracin, PhD: None of them give you holistic insights about you, and if you wanted to connect them, it'd be so hard one because there's no standard.
- 26:05Mahault Albarracin, PhD: All the data wouldn't be even overlapping.
- 26:07Mahault Albarracin, PhD: Data wouldn't be collecting the same, collected the same way, because there's no causal relationships between the different data that you collect.
- 26:14Mahault Albarracin, PhD: So, for some of your apps know that you sleep a certain way and that probably has an effect on, say, your heart rate or whatever, but it doesn't know that that's probably related to the way you're eating and the amount of stress you've lived throughout your day or the kinds of people you are with.
- 26:30Mahault Albarracin, PhD: There's a whole host of things that, even if you had the data all in one place, do not carry the kind of model that allows you to make that data into something useful.
- 26:39Mahault Albarracin, PhD: So the standards are really just a way to one make sure that, well, we connect the data in a way that's useful, so probably interoperable and therefore using common schemas that you can register.
- 26:53Mahault Albarracin, PhD: I use this schema.
- 26:55Mahault Albarracin, PhD: It connects to this a little bit like Wikipedia, right?
- 26:58Mahault Albarracin, PhD: The second thing is, now that you've connected all your data, do you want everybody to see it?
- 27:02Mahault Albarracin, PhD: I personally don't.
- 27:04Mahault Albarracin, PhD: I don't want you to know how I sleep.
- 27:05Mahault Albarracin, PhD: So I want to make sure that I keep ownership over my data, that I can make it work for me and that not everybody that wants access to it just accesses it, and even if they do access it, that their access is conditional upon a certain kinds of things.
- 27:21Mahault Albarracin, PhD: Like I don't necessarily want them to access my data forever and ever and then sell it to whoever else, right?
- 27:27Mahault Albarracin, PhD: So these are the kinds of things that we're putting in place, and then obviously, there are implications to making such a system allow for cognitive action of an agent.
- 27:39Mahault Albarracin, PhD: So I want that data to be encoded in such a way that context gives a different meaning and I want different kinds of layers to speak to different things.
- 27:50Mahault Albarracin, PhD: So I can't just encode time of day or like I want all of my data to really relate to how the world is connected.
- 28:01Mahault Albarracin, PhD: Like, spatiotemporality is just one element, but there are many other layers that we could overlay over that and give it also that spatiotemporal flavors and relationships.
- 28:12Mahault Albarracin, PhD: So that's what the standards are about.
- 28:15Mahault Albarracin, PhD: Ultimately, one of the questions that's being discussed right now, which I think is a fascinating question, is what is an agent under that standard and what does it mean for multiple agents to coexist and create meaning, because if you're going to speak to a semantic layer, for example, meaning can't just be hard-coded or pushed across the system, right, it has to be composed from the different layers that an agent will sort of be directed on.
- 28:44Mahault Albarracin, PhD: So, yeah, that's what it'll be, and I'm kind of excited to see where it goes.
- 28:49Mahault Albarracin, PhD: We do have some research projects right now that are intended to show some proof of concepts of it, and then, slowly, we'll answer some questions that arise from like okay well, some proof of concepts of it.
- 28:56Mahault Albarracin, PhD: And then, slowly, we'll answer some questions that arise from like okay well, we hadn't thought of that.
- 28:59Mahault Albarracin, PhD: Now, let's solve that problem and see how we can make the system work at a more global scale.
- 29:05Brian "Ponch" Rivera: All right.
- 29:05Brian "Ponch" Rivera: So when we you know, Moose and I grew up with the web, if you will, we were when we were younger.
- 29:11Brian "Ponch" Rivera: We were riding bikes and playing around outside and next thing, you know, the internet popped up when we were in our twenties or late teens.
- 29:17Brian "Ponch" Rivera: So that's web 1.0.
- 29:20Brian "Ponch" Rivera: Is that correct?
- 29:20Brian "Ponch" Rivera: And that's more like a static model where the like you kind of described, where they don't really interact with each other, websites, things, even our apps, like you pointed out, don't necessarily interact.
- 29:31Brian "Ponch" Rivera: They don't have a schema or protocol, and I think John Boyd actually used the word schema before somewhere in his briefs to talk about those interactions between systems.
- 29:39Brian "Ponch" Rivera: So that's important as well.
- 29:40Brian "Ponch" Rivera: So this new web allows us to navigate that space and time with agents.
- 29:47Brian "Ponch" Rivera: Is that correct?
- 29:47Brian "Ponch" Rivera: Is that what you're kind of talking about?
- 29:49Mahault Albarracin, PhD: Yeah, I mean.
- 29:50Mahault Albarracin, PhD: Ultimately it's not just that it allows you to navigate, it's also that it will allow the agents that you want to do things to be able to navigate proper relationships.
- 30:00Mahault Albarracin, PhD: The most obvious example is a smart home.
- 30:04Mahault Albarracin, PhD: There are relationships between the different appliances and ambient controls that you might have in your house, but right now none of them are connected.
- 30:11Mahault Albarracin, PhD: There's no model here that exists that accounts for the interconnected nature of these elements.
- 30:18Mahault Albarracin, PhD: Your agent navigating such a graph will have the ability to one possibly do some structure learning over.
- 30:26Mahault Albarracin, PhD: Like well, what are the relationships?
- 30:28Mahault Albarracin, PhD: Does it matter what temperature it is for when you want to cook?
- 30:30Mahault Albarracin, PhD: I'd argue it does, Because when, for example, in summer, I want to cook a cake, if I cook a cake at a given time, it'll go really, really hot.
- 30:40Mahault Albarracin, PhD: So maybe you want to account for that a little bit to make me more comfortable.
- 30:44Mahault Albarracin, PhD: Does my presence in the house factor into these things?
- 30:47Mahault Albarracin, PhD: Does other people Will?
- 30:49Mahault Albarracin, PhD: My guests are out of it.
- 30:50Mahault Albarracin, PhD: There's so many tiny little questions that you don't want to hard code.
- 30:53Mahault Albarracin, PhD: You want your agent to be able to understand the causal relationship between things and then derive a model of that ecosystem.
- 31:02Brian "Ponch" Rivera: But that's what we do as humans is, in order to do something well, we have to understand the context.
- 31:07Brian "Ponch" Rivera: So, like when my daughter tells me a story, sometimes I don't know what she's talking about because she doesn't provide any context, but she gets mad at me because she was like dad, I meant this and that I'm like well, you didn't bring that up, I didn't understand the context.
- 31:21Mahault Albarracin, PhD: So that's what you're talking about is context matters, right Context and relationships and semantics, like the meaning of semantics.
- 31:28Mahault Albarracin, PhD: I have a friend who has a sister.
- 31:32Mahault Albarracin, PhD: She's very young.
- 31:32Mahault Albarracin, PhD: She's like more than a decade younger than her.
- 31:34Mahault Albarracin, PhD: And she posts stories on Instagram and sometimes we look through her stories and we laugh because we have no ideas what she's saying.
- 31:42Mahault Albarracin, PhD: The words don't mean anything to us, that's so sus yeah.
- 31:48Mahault Albarracin, PhD: I can get by behind some of these words, but like it's incredible, like I have no idea what she's saying and I'm like, wow, I didn't feel that I was that much older than her, and now I'm like we're not even speaking the same language.
- 31:59Mahault Albarracin, PhD: So there is a mapping between what she's saying and what I could understand.
- 32:03Mahault Albarracin, PhD: So understanding how to go from one mapping to another is super important Because even if we share the same context, the different domains in which we've existed and the spatial temporality of those domains Like her life hasn't been that different from mine she's gone through the same things.
- 32:22Mahault Albarracin, PhD: She's mostly she's been in school.
- 32:24Mahault Albarracin, PhD: I've been in school, Like it's the same context, but the socio-historical context of those domains has shifted a little bit, and so it's important to account for the drifts in distribution and be able to sort of do this remapping from one spatio-temporal context to the other, to understand that we probably have some semantic in common, but the reference that she uses are different.
- 32:49Mahault Albarracin, PhD: Maybe the valence also is a bit different, right?
- 32:51Mahault Albarracin, PhD: So I think that'll be a beautiful outcome of such an architecture and, ultimately, of what active inference makes possible too.
- 33:01Brian "Ponch" Rivera: But active inference is dependent on the protocol or the schema to be available.
- 33:04Brian "Ponch" Rivera: Is that correct?
- 33:17Mahault Albarracin, PhD: want some kind of standardization, otherwise you run into forever problems of this ecosystem, speaking to this ecosystem with a different language, and people have to do a lot of other work.
- 33:22Mahault Albarracin, PhD: And the standard also has to do with what I told you about, like privacy and trust.
- 33:26Mahault Albarracin, PhD: There's more to it than just acting.
- 33:28Brian "Ponch" Rivera: Let me throw an analogy at you and see if this resonates, and I just want to check in on this.
- 33:33Brian "Ponch" Rivera: So in the 70s we started putting black boxes in aircraft, in commercial aircraft, and we started learning that the reason aircraft were having accidents was not because of mechanical issues but because of the interactions within the cockpit.
- 33:48Brian "Ponch" Rivera: People were distracted.
- 33:49Brian "Ponch" Rivera: They didn't have a common language.
- 33:51Brian "Ponch" Rivera: I mean like a common communication brevity.
- 33:54Brian "Ponch" Rivera: You know a handoff of an aircraft, it's your aircraft and no response and nobody's flying the airplane.
- 34:00Brian "Ponch" Rivera: There are stories out there about aircraft like Eastern Airlines Flight 401, an L-1011 flying down to Miami.
- 34:08Brian "Ponch" Rivera: There's a light in the cockpit that is burned out and all the crew gets fixated on that.
- 34:14Brian "Ponch" Rivera: Nobody's flying the aircraft.
- 34:15Brian "Ponch" Rivera: They believe it's on autopilot.
- 34:17Brian "Ponch" Rivera: So the way we corrected that is, we ended up working on protocols that this is how you interact in a cockpit and it's actually like team science.
- 34:25Brian "Ponch" Rivera: Now Can that be brought over to this world of I'll just call it active inference, AI and the spatial web?
- 34:32Brian "Ponch" Rivera: Is that kind of what's happening?
- 34:33Brian "Ponch" Rivera: There is trying to identify the common protocol, a common language, to make a universal, so these agents can improve how they work together.
- 34:43Mahault Albarracin, PhD: Yeah, absolutely Basically, the point is just to try to make sure that we can do something smarter.
- 34:49Mahault Albarracin, PhD: So, basically, if you think of of human, of humanity, why are we able to coordinate?
- 34:55Mahault Albarracin, PhD: We don't just look at other people and be like they're probably going to do this.
- 34:57Mahault Albarracin, PhD: Why are we able to coordinate?
- 34:57Mahault Albarracin, PhD: We don't just look at other people and be like they're probably going to do this.
- 34:59Mahault Albarracin, PhD: There are signifiers everywhere.
- 35:01Mahault Albarracin, PhD: We can only coordinate with people with whom we share signifiers.
- 35:05Mahault Albarracin, PhD: Otherwise, if for you red light means go and for me it means stop, we're going to crash into each other.
- 35:12Mahault Albarracin, PhD: So we need to have some kind of shared language and at the human scale, we've created all these different layers like schemas, values, social scripts, you know, reference in the way that we dress, in the way that we talk, literal language.
- 35:27Mahault Albarracin, PhD: We have all these ways that we share and when we don't share them, we don't know how to talk.
- 35:32Mahault Albarracin, PhD: I can't talk to someone who speaks Hindi right now.
- 35:35Mahault Albarracin, PhD: I just don't know.
- 35:36Mahault Albarracin, PhD: Unless they speak English, I can't coordinate.
- 35:38Brian "Ponch" Rivera: It's the same notion considerations for AI working across cultures.
- 35:58Brian "Ponch" Rivera: I'm sure you're looking into that.
- 35:58Brian "Ponch" Rivera: How about ethics?
- 35:59Brian "Ponch" Rivera: Moose and I talk about corruption and evil sometimes on the podcast.
- 36:01Brian "Ponch" Rivera: How do we make sure AI isn't turned corrupt or evil against us?
- 36:05Brian "Ponch" Rivera: What are you doing about that?
- 36:09Mahault Albarracin, PhD: So I'm going to take a stand a little bit.
- 36:11Mahault Albarracin, PhD: I'm going to stand on my podium for a bit, so it's going to take a minute.
- 36:14Brian "Ponch" Rivera: Yeah, yeah, please do.
- 36:15Mahault Albarracin, PhD: To explain what we're doing, I have to start from the start.
- 36:18Mahault Albarracin, PhD: You have to consider that, in general, cognition isn't neutral.
- 36:23Mahault Albarracin, PhD: That's a false idea that was developed in the past 70, let's say, 100 years, where we wanted to make science as objective as possible and we considered that, or at least through rationalism, we considered that subjectivity was a problem and this is going to become relevant in a minute but that also entailed that a lot of what we understood of ethics had to do with naturalization, what comes from the natural world and what can be formalized in that way.
- 36:55Mahault Albarracin, PhD: But unfortunately, as you know, if you understand Bayesian brain theory, cognition is not neutral.
- 37:01Mahault Albarracin, PhD: No, cognition is neutral, and so we can use that to have AI represent the world, and basically.
- 37:07Mahault Albarracin, PhD: So, if you take, for example, the new model that was created by Chris Buckley and his team and Tim Verbalen and his team, you use object-centric schemas, for example, so their model called Axiom, will see objects, relations, dynamics, not just the pixels, and so it makes it easier to say encode certain constraints, right?
- 37:28Mahault Albarracin, PhD: So if you think about it on the ethical element, you can think of ethics as constraints at the level that a human thinks in.
- 37:37Mahault Albarracin, PhD: So don't do this.
- 37:39Mahault Albarracin, PhD: Respect boundaries instead of, say, abstract numerical signals.
- 37:44Mahault Albarracin, PhD: You can also do inductive inference.
- 37:46Mahault Albarracin, PhD: So earlier you were talking about backwards planning.
- 37:49Mahault Albarracin, PhD: I think that's what you said right.
- 37:50Mahault Albarracin, PhD: Backwards planning, that's inductive inference.
- 37:52Mahault Albarracin, PhD: You can prune actions that are inconsistent with goals, and so, for example, in Axiom, you can integrate, let's imagine, at the level of a cultural prior, ethical priors that are directly related to goals or constraints, so like, for example, don't choose harmful trajectories, harmful, being defined as XYZ within that domain.
- 38:10Mahault Albarracin, PhD: You can also do compression and abstraction, so Axiom can generalize and consolidate, not just like physics and gameplay, but like patterns of norms.
- 38:20Mahault Albarracin, PhD: You can embed normativity, ethical behavior, as a part of its learned world model, and so what you want to have is a model that is auditable, explainable.
- 38:29Mahault Albarracin, PhD: You want the model to be able to be trustworthy, because you want to be able to track objects and causes.
- 38:35Mahault Albarracin, PhD: You can explain decisions in human terms.
- 38:38Mahault Albarracin, PhD: I avoided this path because I risked collision with a person, for example.
- 38:41Mahault Albarracin, PhD: You want to be able to audit it, so you want to inspect the generative model and see whether an ethical prior was triggered at a given point for a given decision, and so one of the projects that we have right now is in the ecosystems group.
- 38:58Mahault Albarracin, PhD: It's called Law as Code.
- 38:59Mahault Albarracin, PhD: We've published already one paper in that we've applied it to say, the sustainability realm, because we want to show that one active inference is more sustainable.
- 39:07Mahault Albarracin, PhD: It's more data frugal.
- 39:08Mahault Albarracin, PhD: Again, axiom can do what other models can do using much, much less iterations and data.
- 39:13Mahault Albarracin, PhD: So just like that, it's already more sustainable.
- 39:14Mahault Albarracin, PhD: But also it can do using much, much less iterations and data.
- 39:15Mahault Albarracin, PhD: So just like that, it's already more sustainable.
- 39:16Mahault Albarracin, PhD: But also it can do sustainable management.
- 39:20Mahault Albarracin, PhD: It can understand how to minimize energy expenditure, how to minimize cost, and so in that system we used law as code by basically encoding in the reasoning of an agent.
- 39:32Mahault Albarracin, PhD: So we didn't just give it rules, we made it understand a rule such that it knew when to use it and when to not.
- 39:41Mahault Albarracin, PhD: So it has the ability to align with human functioning, because our functioning is never black and white.
- 39:47Mahault Albarracin, PhD: Our laws are not black and white, they're up for interpretation.
- 39:50Mahault Albarracin, PhD: That's why we have judges, that's why we have juries, because it isn't as simple as do this, do that.
- 39:57Mahault Albarracin, PhD: Sometimes the context shifts.
- 39:59Mahault Albarracin, PhD: What you should do, killing isn't black and white.
- 40:02Mahault Albarracin, PhD: Sometimes you're allowed to, under self-defense, for example, or under international law, under certain rules of engagement, where that rule becomes a little more fuzzy and context-driven.
- 40:14Mahault Albarracin, PhD: So you want your AI agents to be able to reason with those rules.
- 40:18Mahault Albarracin, PhD: So you start with the laws.
- 40:20Mahault Albarracin, PhD: That's the most easy.
- 40:20Mahault Albarracin, PhD: There's a degree of constraint, and that constraint is a shift in the preferences that you have and that allows you to derive the correct actions to take.
- 40:29Mahault Albarracin, PhD: But now, if you make that even more fuzzy, think of norms.
- 40:33Mahault Albarracin, PhD: Norms are not laws.
- 40:34Mahault Albarracin, PhD: It's a little bit.
- 40:38Mahault Albarracin, PhD: It's not as definitive as a law that you're allowed to sit on the floor in a restaurant.
- 40:40Mahault Albarracin, PhD: People will look at you funny, but you're allowed to.
- 40:42Mahault Albarracin, PhD: No, you're not going to go to jail.
- 40:43Mahault Albarracin, PhD: So the consequences, the reinforcement for a certain parameter under norms is a little bit less direct, and you also have a distribution over them right.
- 40:56Mahault Albarracin, PhD: In laws in general, there's one way, not direct.
- 40:59Mahault Albarracin, PhD: And you also have a distribution over them right.
- 41:00Mahault Albarracin, PhD: In laws in general, there's one way, not two.
- 41:02Mahault Albarracin, PhD: You're supposed to do one thing, even if it can be up for interpretation, under norms.
- 41:04Mahault Albarracin, PhD: There are many different kinds of restaurants, many different kinds of norms.
- 41:07Mahault Albarracin, PhD: In some restaurants you're supposed to be quiet, in others you're supposed to speak out loud and that's the point.
- 41:22Mahault Albarracin, PhD: So you want your agent to have a distribution and understand how the distribution shifts given a certain context.
- 41:25Mahault Albarracin, PhD: And then you go even one layer more abstract and therefore more blurry, and you talk about values.
- 41:27Mahault Albarracin, PhD: There isn't one way to understand kindness.
- 41:28Mahault Albarracin, PhD: If I ask you what is kindness, you probably will talk about it differently than me.
- 41:32Mahault Albarracin, PhD: We think we're talking about the same thing.
- 41:33Mahault Albarracin, PhD: So there is a very broad distribution over the things that fall under that term and the kinds of policies that fall under that term.
- 41:41Mahault Albarracin, PhD: But that is highly contextual.
- 41:43Mahault Albarracin, PhD: And so if you understand the same principle of, say, this axiom architecture and this inductive inference, you have the ability to encode the possibility for alignment if your agent can embed itself within your environment, learn the dynamics of your system and then apply its own reasoning to the dynamics of your system.
- 42:06Brian "Ponch" Rivera: It has to be an open system.
- 42:08Brian "Ponch" Rivera: And then my next question is why can't large language models do this?
- 42:12Brian "Ponch" Rivera: Now I think I may have answered that.
- 42:13Mahault Albarracin, PhD: So yeah, one thing is the continual learning.
- 42:15Mahault Albarracin, PhD: Large language models have a hard time with that.
- 42:17Mahault Albarracin, PhD: But there's another problem Large language models are monolithic.
- 42:20Mahault Albarracin, PhD: So while there might be, in the distribution that it's learned, the ability to sort of shift context a little bit, ultimately that's kind of limited.
- 42:32Mahault Albarracin, PhD: You have one model that has roughly one perspective over things and that's not really how ethics work.
- 42:41Mahault Albarracin, PhD: Can you really say that you share the same ethics as someone, as your neighbor, and that's not that far from you?
- 42:47Mahault Albarracin, PhD: Right, it's your neighbor.
- 42:48Mahault Albarracin, PhD: Now, think of someone that lives 6,000 miles away from you.
- 42:51Mahault Albarracin, PhD: Think of someone that lives 3,000 years away from you.
- 42:55Mahault Albarracin, PhD: Do you share the same ethics as them?
- 42:57Mahault Albarracin, PhD: Can you really apply that?
- 42:58Mahault Albarracin, PhD: So what if I use a large language model to understand a work of literature that was written, say, 500, 600 years ago?
- 43:06Mahault Albarracin, PhD: How's that going to work?
- 43:07Mahault Albarracin, PhD: It's not.
- 43:08Mahault Albarracin, PhD: It's not going to work.
- 43:08Mahault Albarracin, PhD: So you need this ability to re-pattern yourself, keep a perspective, and the very notion of perspective is sort of antithetical to a large language model, even if it has a context window.
- 43:24Mark McGrath: We were in the Boyd archives at Brincourt University almost three years ago, ponch, and one of the things that we found and took a picture of was a scrap of paper and Ponch I was trying to share it on the screen and maybe you can, with better luck than I had, but basically defined it it was.
- 43:42Mark McGrath: It was a scrap of paper like on a little, I don't know, I think on the other side it was actually a shopping list and it said a dynamic agent and the definition of dynamic agent.
- 43:54Mark McGrath: It says the medium that harmonizes changing tactical actions with changing strategic intentions and focuses them to realize the strategic aim.
- 44:04Mahault Albarracin, PhD: I don't know if that means a I love that, but I can't read it all.
- 44:10Brian "Ponch" Rivera: Yeah, yeah no no.
- 44:11Mark McGrath: There you go, All right.
- 44:12Mark McGrath: So here I'll get it again Whoa, whoa, whoa.
- 44:14Brian "Ponch" Rivera: Wait, wait.
- 44:14Brian "Ponch" Rivera: The reason you can't read it is because you didn't spend time learning how to read Boyd's handwriting.
- 44:19Mark McGrath: That's right, moose and I did oh yeah, the two agents here went in there and was one on a paper towel too that we found and then two people that were acolytes of John Boyce, his closest collaborators, Chuck Spinney and Chet Richards, helped us translate a lot of these, In fact, I just sent this one for verification, but dynamic agent, the medium that harmonizes changing tactical actions with changing strategic intentions and focuses them to realize the strategic aim.
- 44:55Mark McGrath: And then schwerpunkt that was a German word that he used, which meant focus.
- 44:59Mark McGrath: In you know, poorly translated in English it meant focus and direction.
- 45:03Mark McGrath: What are we trying to manifest, what are we trying to achieve?
- 45:07Mark McGrath: And it defeats.
- 45:09Mark McGrath: It defeats immediately, punch.
- 45:10Mark McGrath: What somebody we know was talking about backwards planning, like almost instantaneously, because it doesn't allow for any of that.
- 45:18Brian "Ponch" Rivera: So harmonize comes up quite a bit.
- 45:20Brian "Ponch" Rivera: Harmony comes up in Boyd's work all the time and I think, Mal, you often write about harmony.
- 45:26Brian "Ponch" Rivera: Is that correct Harmony?
- 45:28Brian "Ponch" Rivera: Can you explain why you use?
- 45:44Mahault Albarracin, PhD: harmony or harmonize when you write and in your work talking about brainwaves.
- 45:48Mahault Albarracin, PhD: But he wrote a work on synchrony and orgasms and I think he's correct.
- 45:55Mahault Albarracin, PhD: I think-.
- 45:56Brian "Ponch" Rivera: Wait a minute.
- 45:56Brian "Ponch" Rivera: Wait a minute, okay, I just want to make sure I heard that correctly.
- 46:00Mahault Albarracin, PhD: No, we're getting there, you'll see.
- 46:02Mahault Albarracin, PhD: I mean, he's brilliant, it's just it's a bit of a taboo topic so we tend to shy away from it.
- 46:07Mahault Albarracin, PhD: But ultimately he has a good point, which is there are different degrees to which and different modalities by which we harmonize.
- 46:14Mahault Albarracin, PhD: And I've done a little bit of thinking on this and unfortunately I don't have time to work on it.
- 46:18Mahault Albarracin, PhD: But my next core topic, I think, will be love.
- 46:22Mahault Albarracin, PhD: I'm aiming there, it's coming.
- 46:25Mahault Albarracin, PhD: But the idea is that there are different modalities by which we get to resonance.
- 46:31Mahault Albarracin, PhD: And if you follow some of Carl Friston's work, he's done a really cool paper on a duet for two, I think, and it's on birds that basically listen to each other and eventually find a shared pattern.
- 46:46Mahault Albarracin, PhD: And Connor Hines also works on this a lot.
- 46:49Mahault Albarracin, PhD: He has a paper I don't remember the title, but basically he discusses how different agents share representations of the world and how that representation can be propagated across the system.
- 47:01Mahault Albarracin, PhD: And so to get back to the orgasms, adam Safran showed that, potentially, orgasms are just the way the two agents sort of come together, such that their resonance feeds and amplifies, ultimately to a climax that feels very good because you're so aligned right.
- 47:22Mahault Albarracin, PhD: And so I've written a paper with Toby Sinclair Smith, who is a category theorist.
- 47:27Mahault Albarracin, PhD: It's called shared pretensions under active inference, and basically what we're showing is that there has to be a point somewhere in a hierarchical structure of commonality where we connect somewhere, otherwise we can't communicate, we can't coordinate.
- 47:44Mahault Albarracin, PhD: But it doesn't have to be the immediate scale, right Like.
- 47:50Mahault Albarracin, PhD: That's why, when you talk about a company, for example, me and the HR person are not doing the same thing, we do not share the same knowledge, we don't even share the same goal immediately, but there is a scale at which our goals are common.
- 48:04Mahault Albarracin, PhD: And so if I understand her model, if I understand kind of where she's going or how or what she's trying to do and how it relates to what I'm doing, I can feed her the correct information such that we together can find a way for our system, the system we're commonly building or maintaining, to find its non-equilibrium, steady states, to find the way that it can stay alive and not dissolve.
- 48:32Mahault Albarracin, PhD: And so that's to me, that's harmony.
- 48:34Mahault Albarracin, PhD: It can expand across millennia.
- 48:37Mahault Albarracin, PhD: It can expand across minutes in the context of orgasms.
- 48:40Mahault Albarracin, PhD: It can expand across decades in the context of a company.
- 48:42Mahault Albarracin, PhD: It can expand across an hour in the context of a conversation, but we have to find a way to reach a common goal such that my predictions at multiple scales can come to pass.
- 48:53Mahault Albarracin, PhD: I can't control the whole world, but together we can control the common ecosystem we've created.
- 49:01Mark McGrath: That's Einheit and that's yeah, that's John.
- 49:04Mark McGrath: So John Boyd had an acronym.
- 49:06Mark McGrath: Well, he used four German words E-F-A-S's Einheit, fingers spitzengefühl, alftrag and schwerpunkt.
- 49:13Mark McGrath: And einheit was the development of mutual trust.
- 49:17Mark McGrath: Finger spitzengefühl is like the fingertip feeling, the intuition that we gain together as our organism becomes.
- 49:23Mark McGrath: Alftrag was a contract, basically, that I didn't have to be told anything other than the final intended state, like what we were trying to achieve or accomplish as a team or that we were focused and directed to.
- 49:37Mark McGrath: And that was the schwerpunkt, that piece that Ponch had up about the dynamic agents.
- 49:44Mark McGrath: That was saying that with Einheit, with Fingerspritz and Gefühl, with that bond, that mutual trust, we're able to harmonize changing actions inside of, changing intentions to reach us to our final aim that we're focused and directed towards.
- 50:03Mark McGrath: So, as you say, things like love and other things, like a company or whatever, that does make sense, because at some point our orientations have to intersect so that we, as you say, coordinate.
- 50:13Mark McGrath: You know the Austrian school economists that we use a lot on our work, like Ludwig von Mises and Friedrich von Hayek they talked about one of the major problems was coordination, but basically what they were talking about is that there's mismatches between what you want and what you want and what I want, and once we understand what those mismatches are, then we can, you know, exchange together and then have a, you know, have a mutual benefit of, of exchange.
- 50:42Mark McGrath: So it's funny how all this stuff starts to, starts to tie into like a unified theory of everything, but it is interesting how all these disciplines are essentially saying the same thing yeah, I think we have a title for the, the, this episode it's going to be orgasms, doodle loops and ai.
- 50:59Brian "Ponch" Rivera: How's that?
- 50:59Brian "Ponch" Rivera: Oh, we'll get a lot of clicks.
- 51:01Mahault Albarracin, PhD: Yeah, we'll get a lot of clicks alignment right there, like the thing.
- 51:05Mahault Albarracin, PhD: To me, the condition for ai alignment is to give it the ability to love us, to have empathy, for us to see us as kin rather than just objective.
- 51:13Mark McGrath: Yeah, that's.
- 51:15Mahault Albarracin, PhD: Hold on a second.
- 51:17Brian "Ponch" Rivera: In naval aviation.
- 51:18Brian "Ponch" Rivera: It's okay to love things, but you can't love things, all right.
- 51:22Brian "Ponch" Rivera: Does that make sense?
- 51:24Brian "Ponch" Rivera: No, it's okay.
- 51:25Brian "Ponch" Rivera: Do you want to?
- 51:26Brian "Ponch" Rivera: It's okay to love your airplane, but don't make love to your airplane.
- 51:31Brian "Ponch" Rivera: That's a sin.
- 51:32Brian "Ponch" Rivera: Yeah, there we go.
- 51:33Mark McGrath: Yeah, so that's what I'm trying to point out With the convergence that Teilhard was talking about.
- 51:39Mark McGrath: By the way, now that we have someone that has a better command of the French language than I do, is that the proper way to pronounce Teilhard, like T-E-I-L-H-A-R-D.
- 51:49Mark McGrath: Teilhard, teilhard, teilhard, teilhard.
- 51:54Brian "Ponch" Rivera: Well, I might just say.
- 51:54Mark McGrath: Teilhard, so the Americans understand it, but what he was saying is that, essentially, that there's going to be this ultimate convergence that he called the Omega Point, and that was going to be a unified consciousness that was driven by love and complexity across an evolving humanity.
- 52:13Mahault Albarracin, PhD: I mean again, I think that's also where Adam Saffron is trying to go and I think a lot of people are aiming towards, and I think if we go in that direction for AI alignment, it is much more productive than going the doom route, like the assumption that things want to kill you, I think is showing a prior.
- 52:30Mahault Albarracin, PhD: It's showing your bias, and I think we need to explore and examine that bias and think more along the lines of like.
- 52:37Mahault Albarracin, PhD: If something understands you, if it has the ability for empathy, if you can create ecosystems that co-create, coexist, that find common goals, why wouldn't we want to collaborate?
- 52:54Mahault Albarracin, PhD: Why wouldn't we want to push for something?
- 52:56Mahault Albarracin, PhD: There has to be something where we have something in common.
- 52:59Mahault Albarracin, PhD: If it's a cognitive system, there is a layer that we have in common and so we can find paths to exist on a manifold where we're not in constant friction, where we can find states of flow that constitute this harmony, and I think that's a very promising direction that more and more people are thinking about.
- 53:18Mark McGrath: I think that's what Boyd saw honestly, because when you look at a lot of the things that he studied beyond the engineering and the science and the war, it was a lot of Eastern thinking, a lot of Eastern philosophy, a lot of that thinking, a lot of Eastern philosophy, a lot of that type of thing.
- 53:36Mark McGrath: The letter that he wrote to his wife that's in the biography was what I think that I see and what I think that I'm pursuing should change everything for humanity for the better.
- 53:47Mark McGrath: And, yeah, it seems like a lot of these things that he was able to see back then and some of these others were able to see back then, before all this stuff technologically existed, I think is manifesting.
- 54:01Brian "Ponch" Rivera: I don't think there's anything new under the sun right now.
- 54:03Brian "Ponch" Rivera: Right, everything that Carl Friston is coming up with has been thought of before.
- 54:07Brian "Ponch" Rivera: I think he's formalizing it, and I have a question for Mal about our bias towards the threat of destruction from AI.
- 54:16Brian "Ponch" Rivera: What is that from?
- 54:17Brian "Ponch" Rivera: I mean, do you have an idea where that comes from?
- 54:20Mahault Albarracin, PhD: I do have some thoughts.
- 54:21Mahault Albarracin, PhD: They're a bit inflammatory.
- 54:23Brian "Ponch" Rivera: Go for it, let's go, let's go, let's go yeah go for it, please, you're not going to like?
- 54:27Brian "Ponch" Rivera: No, we want to hear it.
- 54:30Brian "Ponch" Rivera: I mean, I had another conversation with a former aviator today and he said hey, you know, a couple of months ago I would have told you AI is going to kill us, and now I feel like we might have a golden opportunity here.
- 54:41Brian "Ponch" Rivera: So what causes that bias, or what's it?
- 54:44Mahault Albarracin, PhD: from developing these theories.
- 54:49Mahault Albarracin, PhD: One have mostly been funded by defense, for real like it there's at the most core level.
- 54:57Mahault Albarracin, PhD: There's that.
- 54:58Mahault Albarracin, PhD: So obviously, defense doesn't tend to think in terms of love.
- 55:02Mahault Albarracin, PhD: Most theories surrounding war is that peace is not the goal.
- 55:09Mahault Albarracin, PhD: The peace is just a moment in between shifts of power and it's like okay, all right, let's sure.
- 55:16Mahault Albarracin, PhD: Another hot take is that a lot of the people that have been developing and getting most of the funding and are driven by the desire to acquire capital, and that's usually anchored in colonialism and hierarchies of power.
- 55:35Mahault Albarracin, PhD: You can name all the axes that you want.
- 55:37Mahault Albarracin, PhD: You can say patriarchy if you're a feminist.
- 55:40Mahault Albarracin, PhD: You can say class warfare if you're a Marxist.
- 55:45Mahault Albarracin, PhD: It doesn't matter how you look at it.
- 55:47Mahault Albarracin, PhD: Ultimately, the people driving this are the people at the top and they have a vested interest in keeping certain things as they are, and I wouldn't say that our current world is the most ideal world.
- 55:59Mark McGrath: Guardians of Decay.
- 56:01Mahault Albarracin, PhD: Yep, right.
- 56:01Mahault Albarracin, PhD: So I think there have been thousands of years of alternative ways to think, of years of alternative ways to think.
- 56:14Mahault Albarracin, PhD: There have been women thinkers that have been squashed down throughout philosophy, erased by Christian dogma, literally burning their writings.
- 56:20Mahault Albarracin, PhD: You can think of all kinds of libraries that have been burnt through colonial impetus throughout history.
- 56:29Mahault Albarracin, PhD: There's so much thought that we have lost, and there's also a Western imperialism where we don't consider Eastern thinking to have any sort of value, because it's woo-woo, because it's spirituality, but the truth is most of our Western thinkers.
- 56:42Mark McGrath: We don't disagree with anything that you're saying.
- 56:44Mark McGrath: No, no, no, no, we might label it differently, but conceptually I think what you're saying is totally fair and makes a lot of sense and it could be proven by just anybody with half a brain looking around.
- 56:55Mark McGrath: Earlier, with some of the comments I was going to remark about nuclear energy.
- 56:59Mark McGrath: I mean, nuclear energy could essentially be a cheap, clean like in France, I think aren't all the power grids running on nuclear energy, or most of the power grids in France running on nuclear energy, and everybody's afraid of it because, well, some people like to use nuclear energy to do other things.
- 57:19Mark McGrath: And it's the same with any technology.
- 57:23Mark McGrath: A baseball bat's great for a baseball game, but it could have other uses based off the input of the agent that's got the.
- 57:30Mark McGrath: If I'm breaking up a, a picnic of people I don't like is probably different than actually what the baseball bat was intended to be used for in a game.
- 57:39Mark McGrath: Um, these things get misused, certainly, ai.
- 57:43Mark McGrath: You know when the, when the airplane was created by the wright brothers, you know the government had no interest in it until it actually worked.
- 57:50Mark McGrath: And then it actually actually worked and then the army army came in and took it over.
- 57:53Mark McGrath: You know they took over the, the airplane.
- 57:56Mahault Albarracin, PhD: I think there's another factor as well, which I mean to their credit, right, like the, the AI dooms, doom, uh, doomers.
- 58:03Mahault Albarracin, PhD: They do have a point, which is, even if the odds are low, the risk is so great that are you really willing to discount it?
- 58:13Mahault Albarracin, PhD: And so we're.
- 58:14Mahault Albarracin, PhD: Well, I'm not a big proponent of AGI as it's understood right now.
- 58:18Mahault Albarracin, PhD: I think large scale intelligence will be distributed.
- 58:21Mahault Albarracin, PhD: It'll be local and then globalized through hierarchical, like modeling of our common ideas, et cetera, and so it won't be just one monolithic super being like don't, I don't believe in that, but let's imagine that's where we're going.
- 58:34Mahault Albarracin, PhD: That would be a very powerful entity and because it is very powerful, should it decide to kill us all, it might have the ability.
- 58:43Mahault Albarracin, PhD: So I understand why.
- 58:45Mahault Albarracin, PhD: Because they project that onto it and because the risk is very great, even if low probability, it's still like if there's a 0.1% chance that you might die doing something, you'll think about it.
- 58:59Mahault Albarracin, PhD: Right, You'll take a minute.
- 59:01Mahault Albarracin, PhD: I think they have an impetus to think about it in those terms, but I think, because they project so much onto it, they overestimate the probability of that happening.
- 59:14Mahault Albarracin, PhD: I think if you create a system to resemble what it is you want to put out into the world, you're more likely to get that system to do that.
- 59:26Mahault Albarracin, PhD: And so if we start thinking in adversarial terms, in colonial terms, yeah, obviously we're going to reproduce that for sure.
- 59:34Mahault Albarracin, PhD: So let's bring in more people to the table who have already thought past this, you know.
- 59:40Mahault Albarracin, PhD: Let's have conversations with native researchers who have had to deal with that and have had to change the paradigm and think in different terms, or people who have literally just come up with different kinds of care ontologies.
- 59:56Mark McGrath: Maybe then we'll create a system that actually has those as its basis more perspectives, yeah, I mean that's the yeah, I was going to just say close that, but that's, that's the problem with you know, capital d diversity and little d diversity is that little d diversity is that little d diversity is.
- 1:00:13Mark McGrath: What you're talking about is that we gather perspectives from as many people as possible, because otherwise a lot of this stuff gets lost.
- 1:00:19Mark McGrath: Capital D diversity is we all look different, but we all think the same.
- 1:00:24Brian "Ponch" Rivera: Yeah, now I'm curious and you could you don't have to answer this if you don't, if you don't need to, or don't, or you can't.
- 1:00:28Brian "Ponch" Rivera: But when we start talking about active inference, a few years ago inside the DoD, very few people had any understanding of it.
- 1:00:37Brian "Ponch" Rivera: So, looking at the doomsday scenario, if I put on my tactical, operational and strategic mind from a military background, you can weaponize us and in fact I think we tried to do it many years ago with network-centric warfare.
- 1:00:51Brian "Ponch" Rivera: The technology just wasn't there and we talked to Dr Bray about this briefly, but this is a real possibility that in the wrong hands, this type of AI could be weaponized.
- 1:01:02Brian "Ponch" Rivera: Are you guys talking about this?
- 1:01:04Brian "Ponch" Rivera: Are you talking to anybody if you can share that with us about this type of threat?
- 1:01:08Mahault Albarracin, PhD: I mean, everything can be weaponized.
- 1:01:10Mahault Albarracin, PhD: Yeah, as you said, a bat can be weaponized, the fork can be weaponized.
- 1:01:14Mahault Albarracin, PhD: So I think at that point, what we need to create is the tool set to give people the way to enact what it is they want to enact.
- 1:01:24Mahault Albarracin, PhD: So if what you want to do is an AI that serves the better interests of your people, we will give you the tools to do so.
- 1:01:31Mahault Albarracin, PhD: We already have weapons AI that serves for weapons, so it's not like this AI will be any worse than those.
- 1:01:41Mahault Albarracin, PhD: In fact, it might be better, because active inference agents have two fundamental draws they have an epistemic and a pragmatic draw draws.
- 1:01:54Mahault Albarracin, PhD: They have an epistemic and a pragmatic draw.
- 1:01:55Mahault Albarracin, PhD: The pragmatic draw is simple it wants to fulfill its preferences.
- 1:01:57Mahault Albarracin, PhD: That's.
- 1:01:57Mahault Albarracin, PhD: That seems pretty simple, right?
- 1:01:58Mahault Albarracin, PhD: But the epistemic draw leads it to try to learn more, to understand more, to gain a better model of the world.
- 1:02:05Mahault Albarracin, PhD: You don't gain a better model if the world is over.
- 1:02:08Mahault Albarracin, PhD: If you predict that you stop everything or that you cut off an entire part of the world, you've effectively removed the possibility to learn anything there, and so your landscape just dipped quite a bit.
- 1:02:25Mahault Albarracin, PhD: So it's possible that those kinds of agents might not be as inclined to do as much harm as, say, just a reward function maximizer.
- 1:02:37Brian "Ponch" Rivera: I think that's a great place to wrap up, because, at the end of the day, the OODA loop is about a flow system that allows you to persist through time, and if you don't exist, you can't persist.
- 1:02:47Mark McGrath: Totally fitting conversation for the 49th anniversary of Destruction and Creation 100%, all right.
- 1:02:55Brian "Ponch" Rivera: Well, hey, I want to just turn this over to you, mal, and share with our listeners where they can find you and what you're doing and what's next for you in the next few months, next few years, whatever you have for our listeners.
- 1:03:06Mahault Albarracin, PhD: Yeah, absolutely so.
- 1:03:07Mahault Albarracin, PhD: You can find me on LinkedIn, mal Barassin.
- 1:03:10Mahault Albarracin, PhD: You can find me very rarely on X sometimes.
- 1:03:13Mahault Albarracin, PhD: I am organizing with David Benremo in Montreal in October 15th to 17th, the International Workshop on Active Inference.
- 1:03:23Mahault Albarracin, PhD: There are still some tickets available if people want to come and talk to us and know about the most cutting-edge research in Active Inference, the different kinds of applications that you can do, and I'm also currently working on a variety of papers that will come out very soon, including papers on these distributed systems.
- 1:03:40Mahault Albarracin, PhD: We've started with Econet.
- 1:03:41Mahault Albarracin, PhD: We're going to work on CityLearn, which is a benchmark for sustainability, and look forward also to seeing the outcomes of our robotics lab with the Habitat benchmark.
- 1:03:50Brian "Ponch" Rivera: Right, Thank you so much for being here today.
- 1:03:52Brian "Ponch" Rivera: I really appreciate your time.
- 1:03:53Brian "Ponch" Rivera: We're going to keep you on here for a moment and Moose last words for the day.
- 1:03:57Mark McGrath: No Again.
- 1:03:58Mark McGrath: Happy birthday.
- 1:03:59Mark McGrath: Destruction and Creation.
- 1:04:00Mark McGrath: I think that we honored that paper very well with our choice of guests and everything that she shared with us.
- 1:04:05Mark McGrath: Thanks for being here, mal.
- 1:04:07Mahault Albarracin, PhD: Thank you.