 All right. Hello. It is April 22nd, 2024. We're in Active Inference guest stream 80.1 with Laura Desiree de Paolo talking about Active Inference goes to school, the importance of active learning in the age of LLMs. We'll have a presentation, then some discussion. So thank you, Laura for joining. Looking forward to this. Thank you very much for having invited me. Okay, so I'm going to present practically one of our latest publication, which is in fact that the inference goes to school. Oh wait, I have to. Yes. This is an overview of the talk more or less. I'll try to go as fast as I can for the first part, because I'd like to spend a little bit more time on the later slides. And I'm sorry for some misspelling or mistakes that you will find here and there in the images, because I had the long fights with church APT for trying to generate images that would vaculate better my talk. And at some point after 50 or 60 attempts I simply gave up and they choose just the best one that I could achieve. Okay, so I still have to use my mouse. Just few things about myself. I've been interested in learning since I started university, maybe because I was considered a slow learner in high school, definitely not a high achiever. And therefore when I started to study at university I started to study learning, different kind of learning, different learning across species and the evolution of learning and so on and so forth. Till the point in which I started to be interested in learning in educational settings thinking, okay, this is the perfect recipe for success, because you have all the ingredients that you might have. You have super curious beings for who's ever had to deal with children before they go to school, knows that they have this insane curiosity and they ask why is about everything. Theoretically in school you should have tons of learning opportunities and you finally have people that are trained to give the answers that the children seek to have. Unfortunately the situation in school is not always so nice for quoting the principle, the British principle during an interview who said it's incredible that children prefer to be home sick than going to school, the place in which they can learn stuff. Consider the learning is fun and for what concerns me I am a researcher and as I assume most of us, I love learning, I think learning very, very fun. So I feel part of my job or my duty to bring back the fun into learning, particularly in formal educational settings or in compulsory education. This is just a slide with some basic literature about active inference and learning. I won't be too specific about active inference. There is here a veil that eventually could follow up for some questions more specific about the active inference because it's way more prepared than I am. But anyway, this is like, I don't know if you have, I have to move this one. Okay, perfect. This is a basic model of active inference, the idea of cognition in active inference for which the agent and the cognitive system of the agent is not waiting for the information to arrive and to fill in. But instead act proactively, so make predictions about the state of the world in which I expect to leave the information that I expect to perceive into the world. And then every time, compare the predictions that has made with the effective information, the effective situation of the world in that particular moment. And when there is a discrepancy discrepancy in terms of both surprise or events that are unexpected or errors, either update the model for making better predictions or change the words and so act on to the word for fitting the original predictions. And then this model of cognition learning is a sort of imperative because every time that there is a discrepancy between the prediction and the state of the word is like a red flag that says, okay, here there is something that can be learned that there is something new that can can be acquired. For making my future prediction my future and error minimization better. Another interesting aspect of the active inference is that predictions are not just in terms of thoughts, beliefs or knowledge, for instance, but also in terms of sense of motor activities. So everything that we perceive, feel, see is considered a prediction about the word and it's considered in itself cognitive, a way to understand the word that surrounds us. The model of active inference of learning and cognition together is very efficient and is very powerful for explaining, for instance, different kinds of learning or learning mechanism, using the same very simple, so to say, mechanism. You can explain, for instance, individual trial and error and the imitation as every time the generative model that optimize on the environment and on the situation environment for acquiring the right information or for making the best predictions possible. You can also explain how expertise in terms of refined knowledge, what is called know what or in terms of skills can very easily come from explorative policies and for curious behavior into the world. Also, importantly, focuses on the role of attention that if you think in terms of educational settings, educational environment is, of course, very important. Again, in terms of prediction, aerodynamics and surprise for the position of competencies and skills. And most importantly, at least for me, using the TV inference to process that very often are considered complementary, but separated, which are development and learning. They can be recombined together because making predictions, particularly in terms of sense of motor activity and sense of motor interaction with the word means also acting on our biology and neurobiological structure for being sure that these predictions are as accurate as they can be. Okay, when we wrote these, when we wrote the paper, we had three main problems. The first problem was how do we apply a theory of cognition such as the active inference in real educational practices and conversely how these educational practices should look like for fitting our theory of cognition. On these two problems adds up a third problem, which is the use and the spread of artificial intelligences, particularly large language models in education, as well as in other fields, and how these might change the rules. The problem of application of large language models, particularly artificial intelligence or even other form of technologies within educational settings, but particularly large language models, is given mainly by the fact that they are disembodied. When a user has to deal with large language models, in reality, deals with some artifacts, the computer, the screen, the keyboard and so on, that can touch, see and perceive, but at the end of the day deals only with the input output series. And the whole process, all the other artifacts that are printed, that are inside the model itself are completely invisible. Which makes a problem because normally we are used, for instance, we can't make behaviour reading or understanding intentionality behind the answers of ChaGPT. Now I added yesterday morning this slide, this is a beautiful paper that came out, I think, two or three days ago, about children, the perception of children about information on Google. And it's very interesting because somehow it has the same problem that LLMs have. Children, even very young, children trust the information that Google gives, but they don't have any possibility to understand where this trust comes from. In the moment in which they have to deal with something like ChaGPT, in which the answer is given in a sort of natural language, of course the problem is even deeper. Therefore, how do we nurture the human generative model in the age of the artificial intelligences? These pictures that I have here on my slide, I was looking around for different representations of how integrate artificial intelligences within educational settings. And they all have one thing in common, which is they tend to treat the artificial intelligence in a sort of human way. Because this is if the artificial intelligences are disembodied, it states that we are embodied, embedded, extended and enacted agents. So we are active inference agents. So our generative model for understanding, for dealing with disembodied entities has somehow to put them in a way that we can understand. So our problem would be how do we build an educational environment in which our generative models can effectively and efficaciously deal with the powerfulness of these new technologies. One way is what I call here the traditional way. This is the only image that I got from ChaGPT that was perfect at the first attempt. I asked ChaGPT to give me the image of a classroom with pupils between 8 and 10 years of age. The fact that ChaGPT gave me directly, this image tells me a lot because ChaGPT is trained on our human knowledge. And this is very likely the image that we all have when we think about a classroom about education. This that I called here traditional educational setting in reality as a specific name, which is the teacher-centered approach, which I have somehow schematized here. And more or less works this way. You have all the human knowledge that is far back in the background of the educational setting itself. This extensive human knowledge is parcelized and organized in curricula that then affect the child. On the other hand, you have a teacher that also affects the child. And you have this child that somehow acquires this information and grows up having been filled by the information that is provided with. Most scaringly, this kind of structure is very similar to the kind of training that we do for disembodied entities such as LLMs. With the additional problem that defined tuning in educational setting in these kind of structures happens at different levels. Not only the child is fine tuned by the teacher and by the curricula, having somehow knowledge that exists, other knowledge outside of the classroom, outside the knowledge that is provided within the educational settings. But the fine tuning works on the teachers themselves and the teachers themselves, they don't fine tune every single child, but they usually fine tune our whole classroom. Now, for whosoever knows how the fine tuning for instance in LLMs works, knows that the more you fine tune a model, the least creative the model is going to be. Now, when we think about these kind of traditional settings and we think that the activity inference is the best model of cognition that we currently have, we can't avoid to recognize that it works also in these settings. Problem is, yes, it works, but it's not really on point. In the sense that, for instance, very often in these kind of settings, we pretend that the attention of the pupils goes specifically to certain feature of the environment, the lessons that are provided by the teachers. While, in fact, we know also that children at particularly in early age, so till around about 12, they tend to pay attention way more and to learn a lot from irrelevant contextual information, which in active inference terms makes a lot of sense because it is a generative model that had till that point in time very few experiences. So it's gathering information in the environment for making a better prediction on the long term. Another problem with this model and this is the only concept that is not in the paper that I have added in the slide is what I called here error learning dynamics. Another problem with these kind of settings is what we consider the error signal. We have said before, in the moment in which there is a discrepancy between the prediction that a cognitive system has made and the actual environmental situation, there is an error signal. If the error signal doesn't come exactly from the mistake but comes, for instance, from the feedback that might arrive in terms of notes or from being scolded at home because you get a bad grade or you got a note in school, then of course the change of the situation, the acting onto the word and the update of the generative model and of the prediction doesn't fit with what we want the children to learn. These studies that I have here quoted, some of them are really beautiful studies of Solange de Neveau. She has observed children in different educational settings and also children and adults how they react to mistakes and how they monitor mistakes and discovered that children educated in this kind of traditional or teacher-centered approach educational settings develop more than the ability to learn from mistakes, develop what is called error avoidance, which again makes sense. It's not that the active inference model doesn't work in this way. It's only that the error signal doesn't come from the mistakes, for instance in the task or in the exercise that the child has made but comes from an over-aspects of their word. Okay, therefore how should the classrooms in which we want effectively to foster an idea of an active inference agent to nurture an active inference agent? It has to be an educational setting, it has to be a classroom that offers explorative opportunities and offers the time for precision-weighting. What I mean with that? It has to be an educational environment in which there is the possibility to learn from various aspects of the word and there has to be the possibility for the child, for the pupil to fine-tune its generative model multiple times on the same task in a way that can really understand at best how to perform it. It has to be an environment in which it is fostered also a multi-modal and multi-sensory way of approaching learning and knowledge because as we said the predictions are not just in terms of know what the information that is provided to the child but also all the irrelevant stimuli that are around, so irrelevant, so they shouldn't be irrelevant, should be instead relevant. And as to foster also individual interests and attention modalities because at the end of the day the generative model is always individual, it's always the individual that updates its own generative model in its own specific environmental settings although of course some coach forces as Abel would call them that makes us being able to act and to think in terms of our community, of our culture. It has to be an educational environment in which developmental learning are considered together and not separated. In these traditional educational settings, for instance, we have all the classroom that very often are divided by age, not about skills or stage of development. It has to be an environment in which the construction of abstraction and categorization has to come through environmental experiences which again in active inference terms makes a lot of sense, constructing abstract knowledge, constructing abstract categories allows the system to be somehow prepared to how to deal with certain subject or with certain object or with certain circumstances and then requiring just the fine tuning of the generative model on the specificity of a particular circumstance. And it has to be an environment in which the error learning dynamics are considered in a way that the error signal comes directly from the error and not from other contextual information. And instead of thinking to invent from the top of our mind a new educational approach, we looked around to search for which educational approach that's already out there that offered the most opportunity to fit with our idea of educational settings. Amongst various, the Montessori method is the one that if doesn't fit all the boxes, definitely fits most of them. For describing a little in the Montessori method, you will always find a sort of open spaces and multi-age environment of learning. The curricula are always found on three to six years and not on just one year as in traditional schools. Children are always free to choose what to study for how long we fume. It is already implemented within the pedagogy and within the theory also of Maria Montessori, the idea that learning and development are just one process. It works through what is called the indirect form of teaching and through the prepared environment. The prepared environment is the classroom as it's called in the Montessori method, which is already organized in a way that can foster some kind of not only of skills but also of construction of specific knowledge and pays a lot of attention to the error dynamics through the design of the classroom itself, but also through the design of the specific learning material. I have some examples in a couple of slides. The materials in a way that the material is designed in a way that a pupil, particularly in the lower stages, so in the first grades, can immediately perceive when they have made a mistake and can immediately correct them. So the error signal arrives directly, it doesn't, there is no delay. And the whole classroom is organized in a way that the uncertainty, so the prediction or minimization is somehow manageable and still organized in progressive difficulties, but every time adding up just one extra variable, one extra element of uncertainty. Now, it's well known, if anyone wants to know more about the history of Maya Montessori, I can talk about it later, but it's well known that the method was invented by Maya Montessori, an Italian physician at the beginning of last century. I have been lately dealing with the most applied of her work, particularly in mathematics, in geometry and in grammar, and I have to admit that she was frankly genial. The design of her material is astonishing still today. The Montessori method is incredibly famous and incredibly well known. There are more than 15,000 Montessori schools in the world, in more than 140 countries. In 2011 came out a paper in business psychology, I think it was called the Montessori mafia because some of the richest and most well known entrepreneur in the world are aware Montessori pupils. But still, it's not so much well known, the theories of development and learning of the Montessori method are not so well known, for instance, in cognitive sciences. So they are very much, I was very surprised the other day looking at the how many results one can have putting Montessori and the entrepreneur in Google scholars and not so much in cognitive sciences. This is not really related to the topic but is also a good aspect of the Montessori method, which is given the particular attention that gives to the ecological context for learning. It has also a leading position in terms of ecological sustainability within the educational setting and within the curricula. And this is, of course, a topic of major interest that should, in my opinion, at least be part of whatever curricula in whatever school independently of anything else. Okay, these are some examples of Montessori schools. The Montessori schools are more or less all structured in similar ways. First and foremost, there is no distinction between inside and outside the classroom. Every space given the physical constraints of the buildings might be corridors, might be small or big rooms. The doors are always open, the children and the pupils are always free to move around. There has to be always a free, a big free space available for children that prefer to work on the floor. And also because some of the materials requires lots of space, there is the material for history. The history band that's up to five meters long and is designed exactly for the children to understand, to have a perceptual understanding of the different positioning in time of the different characters or events that they are working on. The inside and outside, as I said, they are both educational in two sense, not only because there are lots of activities like the observational nature and part of the curriculum of botanics, biology and so on that are literally run outside. But because the old furniture and the old everyday items are easily portable outside and very often you might find the classrooms or the normal daily work happening in the outside space. There is one of the images that I have here in black and white taken from the book of Schneider and Müller refers to a Montessori school in Germany in the 30s. So it's historically designed in this way. We find the Montessori classroom also the construction of abstract categories and of categories in general is supplied by the environment itself through the positioning of the objects of the materials, the books that refer to specific disciplines of specific subjects that are positioned in a way to construct more and more abstract categories. For instance, you might find the materials for botany and the materials for chemistry and the materials for physics all under the label science and each one of them will be labeled as well. In a way that children simply interacting with the space, interacting with the material and simply with the simple activity of taking the material out from the shelf and put it in back, understand more or less how the different subjects are related to one another. Plus, for instance, through the presence of kitchens that I used for making food, I used for experimenting chemistry, I used for taking care of plants and I used for washing the brushes, allows kids to make connections across different disciplines. When I talked before about error learning dynamics and how the error signal should immediately get back to the system for having an effective update of the model or modification of the word in terms of effective learning. This is one example. This is actually the first material for mathematics that is in the Montessori method that's usually presented to children between the age of three or four. In Italian, they are called the numerical roads. In English, they are called the red and blue roads and these long pieces of wood divided in a section that represent units, divided in different colors, so in alternative colors, blue and red and they are from one till ten, with the ten that's one meter long and the one that's ten centimetres long. These roads have the aim to help the children to learn the basics of addition, addictions and subtractions, making perceptual evidence how the different quantities are related to one another. Most importantly, they don't have the aim to teach the quantities because children of three or four age, they already ask for three bonbons instead of two, so they already have an idea of quantities. This kind of material is designed exactly for making clear how quantities are related to symbols and how quantities are related to one another. This is the first of a set of three materials for the beginning of the curriculum of mathematics, which as a matter of fact has this representational aim. The other two instead that are the spindle box and the numerical coin, the spindle box has instead like sort of the opposite aim. In this case, the child has in front of them a box with already some, the series of numbers, numbers impressed on the box and with 45 spindles, all identical, 45, 9 plus 8 plus 7, etc. And the work consists in taking the spindles out from the box and putting them back in a way that the number of spindles in each compartment has to be the same that is represented by the symbol itself. As we see here, the series of number goes from zero till nine. In this particular exercise compared to the other one, we will have just one extra element, which is the child has to understand that in a series of numbers, zero represents no quantities. So it has the same, so it somehow reinforces the knowledge that has previously acquired with the other exercises and now adds up with an over element. The third exercise in mathematics is the numerical coin. In this case, it works as sort of test for children of three or four years of age in which the children have not only to combine the symbols with the quantities but also to put the symbols in the right order. Of course, there is no shame in going back and in doing the different exercise at different levels. And this is very important because as we see here, I'm sorry that I used Italian, but after having had the very bad experiences with the English translations, I now tend to use directly the Italian. This page refers actually to the different calculations that can be done using this kind of material at different stage of development. If a child of three or four might do the basic arithmetic of nine plus one, eight plus two, et cetera, noticing how nine plus one, so combining the road of nine plus the road of one, you have the same measure, the same dimension that the road of ten, so it's equal ten. A child a little bit older that had already some experience with other forms of calculation, for instance, multiplications might notice that combining the roads, you might have five roads all with the same length, the length of the ten roads and you will have left the five roads in this way making a different calculation. So calculating the total number, the total quantity, the total number of units in this series, which is ten multiplied five plus five, which at the end gives the possibility at, again, a different stage to calculate, so it gives the basics for calculating the complete quantity within a series of given numbers. Okay, and I swear I am about to finish. These are three of the three other elements of the extensive curriculum in mathematics and science. It's not a case that the Montessori method seems to have the best result in mathematics and science because the curriculum, the materials for mathematics and science, particularly for the first grade, is the most extensive one. Now I'm not describing in detail all this material, just the pearls that are the one on the left of the screen are used for making a little bit more complex calculation and also for visually teaching the pupils how to calculate cubes, volumes. But some times ago I attended to this beautiful talk of Antonio Torri that is also an implementation in school for the study of computer sciences in primary education. Through a concept that is called the concreteness fading. So these are implementation for understanding how to help pupils, particularly young children, to construct abstract knowledge, particularly for very abstract competencies such as computer science. The concreteness fading approach practically consists always in three moments for which we start from a concrete element, in this case it's the dead man island, it's a sort of game that the kids have to perform. An intermediate phase in which you add to some elements of the concrete phase, some elements that will be in the more abstract phase for then finally end up with the abstraction. This kind of concreteness fading approach is everywhere in a Montessori classroom. And it has the advantage that is not just an implementation because the implementations are usable and beautiful. The problem is that once the implementation is brought away, we have the classroom as it was before. Instead, in this case, these kind of elements are present in the classroom at all times. And this also helps for somehow solving some problems with this approach for helping children in constructing abstract knowledge, which is as always in science, in particular in literature education. Some people think that the best way to help kids in constructing abstract knowledge is going from concrete to abstract. Some people think exactly the opposite to present first the abstract knowledge and then to allow the kids to make connection between this abstract knowledge to the concrete elements in everyday world. We can assume very easily using, again, an active inference point of view that both approaches might be right, very much depends on the kids. Having these physical elements all the time into the environment allow different kids with different generative models to construct their own knowledge in their own way. Some kids might find easy to go from concrete to abstract. Some kids might find easy to go from abstract to concrete. Some kids might find easy to go from intermediate to abstract to concrete. But anyway, they will always be surrounded by the same objects for extensive periods of time. And this allows also to refine and to maybe going from one to the other approach at different stages of development. Now, the Montessori method, of course, is not the only valuable method of education that exists in the world. This is one of the ones that I like the most. It's called Learning by the Sign. It has different voices in educational literature. This is just one of them. And I particularly like because things to change or to improve educational settings, not from the head, but from the feet. So not through educational pedagogies, but really and physically rebuilding the environments in which the children or the pupils learn every day. The problem with this and other approaches of this kind is that they are not particularly precise, in my opinion, in looking at the error learning dynamics. And if we don't pay attention to what a pupil perceives as an error, what gives the error signal to the generative models, of course, the learning is way more difficult and is way more complicated. At the end of the day, when we want to design educational environment, we want any way to transmit certain kind of cultural and cultural competencies and cultural knowledge. Finally, we also haven't dedicated much at the end, much thoughts about LLMs. The point with LLMs is that it's clear that they are here to stay and is somehow recognized that they will have a huge impact on education. The problem is how do we implement them? If we implement them in a sort of traditional educational settings, they will do for the pupils what teachers do, which means they will provide knowledge. But we know that the knowledge that is provided by LLMs is not always accurate, is sometimes messed up, sometimes it's mixed up, sometimes it's clearly invented. Therefore, it's if we want to really use them within educational settings and we should, because they are powerful entities, we should do that with what Kasneici et al. defined a clear educational approach, pedagogical approach, with a strong focus on critical thinking and strategies, something as the Montessori needed. Finally, the idea that I have that we have is more or less to combine the elements, some elements that are taken from the Montessori, from the Montessori method, particularly for what concerns the application in educational settings. Because this method already has in focus not only the child, but also the teacher and also the curricula, which is always somehow regenerated by the interaction between the child and the teacher. I think maybe I should have mentioned earlier, in the Montessori method, the interaction between teacher and pupils is seen as a form of collaboration, not as a form of giving some kind of information. And also that, as in the focus, in the center of the focus, at least part of the extensive human knowledge that we have created in Millenia, compared with, combined with an approach of learning to learning and to cognition, which is given by the active inference, which is a predictive approach to the relationship between a cognitive system and its environment, which, in my opinion, at least, might be really the game changer in education. When we will stop to think about school as a place in which the children and the pupils are filled with information, but a place in which they will try out what they already know, and they will refine their knowledge. This will be really, I think, a game changer in educational settings. And that's all. Let me just thank Avel, Valgenin Karloot, Ben White, Axel Constant, Andy Clark, of course, that are part of my team at the University of Sussex. The chat lab, the Children and Technology Lab, Seattle University of Sussex with the wonderful Nicola Yule, and the Escape Material Minds a Year C grant that pays my salary and allows me to work on different aspects of materiality, for which I can work from educational aesthetics to making activities in early home, and this is really fun. Thank you very much. Thank you, Laura. Awesome talk. Cool. All right. I can stop sharing this screen here. All right. Thank you. To kick things off, Avel, if you want to give any first reflections and questions, and then meanwhile, if people have any questions in live chat, I'll look for that. Yes. Thank you, Daniel. So thank you, Laura, for the talk. It's clear. It works. Good thing to having a talk. Something I wanted to say in addition, for instance, with what you said is that you were mapping the re-correct point out that an easy way to do things would be to say that active inference tells us that we need to learn in a certain active way. That is not what is happening. Active inference is a model of cognition. If it's true, cognition is predictive. And whatever we can do will not change that. So active inference does not tell you that it's better or that the natural system is unnatural. That is not something that happens. But what you did and what I'd like to dig in the general sign intuition, but in a different formwork way of thinking, is that while it does not tell you that one way to learn is better, it does tell you that while learning is contextual. If you learn to predict things, you learn to predict specific things from specific stimuli. So the way that you will learn things will educate what type of stimuli you get attention to and what type of feeling you give to. And the last 20, 40, 100 years, they were characterized by a very, very strong cultural selection pressure for institutions that could control people doing things at a very large scale. So mostly nation states, mostly with war. And so they have a very strong interest to for people to understand things through the, basically, what the higher up things of it. And it happens that application system trends you to worry not about what it is that you're doing, but what the higher up things you should think about what's doing. And if we are, if we are lucky, possibly we may have what's already occurring, I guess, well, that is more and more centered around like problem solving. We are a city, we need to do something with your, what should we do? And this is an open question. And your higher up is not like simultaneously a sociologist, waste treatment, major physicist, etc. So you'd have to have some kind of active the march of trying to find out what it is that we should do. And the Montessori does that. It does teach people to like, okay, you have a problem here is the environment with the problem is set. Go ahead. Do what you want. And these trends people that are, I guess, I hope more efficient in the problem solving that people who are basically trying to learn the formula and repeat it and then perform as a repeating formula. So I think there is a pretty deep thing to be that. And I think you've identified it. And I think that there is like it's interesting to think of it in terms of active inference and there is more way to go in understanding what precisely in terms of active inference. Sorry, in there is more work to do in understanding in terms of active inference how precisely Montessori or classical school system shapes the way that people think and organize. Can I answer? Yep. I completely agree with you, obviously. But and maybe I wasn't precise enough. So the point is that, as you said, active inference, if this is the way in which this is the way in which our cognitive system works, it will work in the Montessori method in the Waldorf schule in whatever system it will be. The problem is how much in the moment in which our intent is educative, how much we are precise in thinking the environment and the precise context of learning for being sure that the pupils will acquire the specific information that we want them to acquire, which doesn't mean that we want all the kids to think the same. But we want, for instance, to transmit a certain level of basic competencies, basic mathematics, basic literacy, right? From which then they can build up by themselves. For doing that, I called them here because I didn't have a better term ever learning dynamics. Because the point is, as you said, when a child goes to a traditional school or whatever, I am personally against grades and notes and all this kind of stuff because there is a bunch of literature that say how bad they are for motivation of the kids. But for me, they have also an oversight effect. First, they are delayed in time, which means if we think in real active inference, there are posteriors that never refer to what they should refer to. If I have a bad grade for a test that I've done last week, I receive my error signal doesn't come from the test. Doesn't come from the mistakes that I've made in the test. It comes from the red notes that I received from the teacher. What I will try to fix in my environment and how I will try to update my generative model is going to be, oh, I want to avoid that. This is my prediction error, my attempt on the long run to solve uncertainty, right? But this doesn't fit with the idea that we want the child or the pupils or whatever or even the young adults to learn the content that was in the test in the first place. This is, in my opinion, the main big problem. The second problem, which somehow the Montessori solved, but the active inference is better, is how we think, how we have to design learning. Because if you think learning as something that feels in the brain of the child, thinking in active inference terms is, in my opinion, even more precise at thinking in terms of construction of knowledge, right? Because the knowledge is already somehow constructed. You arrive and you try it out, the knowledge that you have. And then you refine and refine and refine. Is this approach that I think could help in designing learning environments, learning materials, how we design the books, how we design the interface, how we design the team group. They have to be designed in this way. A child that thinks about the basics. I try to think in very, very practical terms. A child that arrives to school is six and that's to learn to read and write. And the first thing, this is the Montessori myth on this, it's okay because theoretically you teach literacy at a very early age. So you make the children play around with letters at the age of three or four. They literally play around with big letters. But theoretically, when you go to school and you want to teach children to read and write, what do you do? The children arrive the first day in class and they start with the A. Tell me, how can a cognitive system predict something as such when a child of six knows how many words are connected by the thing that you want to learn? This is complete nonsense, in my opinion. It's complete nonsense. Instead, if you think the other way around, what does the child know already? The child knows stories, the child know characters, the child knows structures, the child knows context. And you think to use these as form of predictions that the child will make. And every time you add one single element of uncertainty, manageable uncertainty, then it's going to be way more natural. I'm sorry to say, but that's how it is. Okay, I'll read a question next from the live chat. Galea Maximilist wrote, what is the Montessori perspective on peer learning and social clicks in the classroom? How does this relate to active inference modeling of learning? The Montessori approach to learning is a little bit complex. So in the Montessori approach, learning and development, they go together. And you have the phases of development, which in my opinion fits way better with the active inference that the tier of development or Piaget or Bigoski. They are kind of rough periods of development in which you construct abstract competencies. But if you think in specific terms Montessori think that the first competence that the cognitive system construct is the capacity of absorbing from the world. Okay, this is the first thing that a child does. Montessori describes the infant, the newborn as searching into the environment the information that he can get. It has very much to do with our approach in active inference terms of the model of learning. Albeit, of course, is not so precise. First, for historical reasons, she didn't have the right terminology. She didn't have the right knowledge. She didn't have the right neuroscientific knowledge. Many of the things that Montessori said or claimed they have been proved only in the past 50 years or so. There is a beautiful book on Montessori neuroscience written by Leonardo Pogazzi, the guy of the mirror neuron, one of the guys of the mirror neurons. And for instance, about the role of the hands in cognition and imitation, the role of the activities, the role of perception. They are already all of them inside the Montessori method and they very much are in line with our approach, our active inference approach to learning and cognition. I'd say that they are compatible. I should have started with that. They are compatible. I hope that answers. If it doesn't, keep going. Yeah, well, you made a very provocative comparison with how the LLMs are trained with a corpus and a little bit of fine tuning. And then basically the content delivery instructionalist approach where there's a corpus, there's the textbooks and the material. And with some fine tuning, we get to what a corpus regurgitating machine. And along with that, there's also these kind of implicit cultural learnings. Not just related to the embodiment and the layout in the classroom, but things like I should be predicting what everyone else predicts or the next piece of information on my foraging journey will be provided to me. Or there is an expertise hierarchy that's inviolable and all these different kinds of higher order properties that that become learned through the education system. And I think Avel made a great point, which is that active inference is more of the tool that helps us analyze these different cognitive settings. And then talk about how they align or not or how they fulfill our preferences or not. It's just, it's very interesting to use this new lens to bring together many of the critiques that people have raised about education. And I guess how do we move from taking the lens and applying it in a critical fashion to proactively designing new curricula and new experiences for for younger and for older individuals? I remember at some point, I don't I don't remember whom, but it was a critic of education during the talk is that there is just one way to to reform education. We have just to like to cancel schools. That is the best case scenario. I'm not so extreme, but the idea to change. The idea to bring something like the inference within educational settings means, first and foremost, not thinking about education in terms of curricula. Not thinking about education in terms of what students should know at the end of the year or what teachers should teach in one year. But is is thinking about environment in which you foster options. Thinking about in active inference terms means in active inference cognitive system. What does particularly are new being cooking TV stands. Active inference agent. Try looks around simply looks around for gathering information because they might be useful for making prediction error, accurate prediction errors in the future. Which, how do you translate this into location? Means offering opportunities. Means offering as much as it possible for making better prediction errors. Right. Means shifting completely the mindset. From what a person, a child, a pupil should know to what we can offer. What they can know. It's clear. That is for me, it's the mindset that I'd like to shift. Yeah, that's very important like thinking about the adjacent possible and the perception and the affordances of each individual. And that's also how education and accessibility can come together, rather than trying to take like a top down organizational approach or still again a content delivery approach, just slower or diluted rather than actually like welcoming people's differences. Um, okay, glia maximalist rights. What about the social structure of the classroom? First question. And then, okay, go for that. And then also second question. No, no. Okay, social structure and. And do you think that grades are beneficial for older children? It reinforces that letter grades are the delayed readout of performance like many aspects of adult life. Maybe we should take off the grades also from adults life. That would be way easier. Social structure in in the class in the Montessori class on the social structure are very much again natural. So you have the kids that forms groups. They can always work with whosoever they want. So they form sort of natural friendships. So you work more with the people with whom you work better. So to say. And the social structure, of course, I couldn't go too much into detail about different social learning. But if you think how the information is provided in a Montessori classroom, the teacher never teaches. So you will never find a teacher that tells you some stuff and you have to repeat this stuff. Most of the time, particularly for basic competencies, the teacher shows how to use certain materials or helps you in finding information online. Things as such. These kind of approaches and Montessori was incredibly precise. And really sometimes when I read her, really like I'm astonished. It was incredibly precise, saying how the teacher should talk with with young children using gestures, using certain gestures in a way that they could better vaculate the message. If I think in terms of social learning, and I think about five or four years old, six years old, and the teacher, which is a adduced, competent individual. And I think in terms of social learning, what does it do? Stimulates over imitation, which is what is considered the typical, the most common of our way to acquire information socially, which is also in traditional schooling setting, the one that we never use. Have you ever tried to copy it during a test? It's the first, the first thing that you're supposed not to do, but is the most natural way in which we transmit cultural information. For saying about the social structure, this kind of different mechanisms that again, according to the active inference, I see, I see lots of advantages to use active inferences for learning, because it allows to think the different kind of modalities in which we learn, we transmit and we learn information. Cultural information among each other as a society, not as separated in essence, but more like a sort of spectrum. I try alone, if I don't succeed, then I will look, I will shift my attention to something else and so on and so forth. In a Montessori classroom, this is allowed. Children can try out by themselves, if it doesn't work, they can work collaboratively, collaboration. If I want to talk about not only social structure, but also the social learning mechanisms that are involved, they work collaboratively. They imitate each other all the time. The imitation is not simply used like in traditional high school, you imitate the cool kid in the corridor for how he's dressed up. You do the same in a Montessori classroom, just so you can work with them. And usually it's not the child that is dressed the better, but the child with whom you work better, that is your cool child in a classroom. This changes a lot the dynamics. Grades, grades, grades, grades. Right now there is a big movement against grades. I think they've been published at least four books in the past five, six years against grades, the use of grades in education. For older kids, the problem with grades is that there is a gap between whatever system we have for children in compulsory school and when we go to university. That is the main gap. Till the university works with exam and grades and so on and so forth, we will have to keep the grades into the high schools because university allows the access based on some numbers that are given to you. The later we put the grades, the better. And the reason is very simple. Very young children, they haven't yet developed the ability to understand, to make a difference between what they do and who they are. I am a bad note, I am a bad person. It's very hard for them to make all these levels. These are all kind of refinements of the generative model that comes when you have multiple experiences for which when you are 17, 18, 19, you start to discover that you might be very bad at doing something, but you might be very good at doing something else. For younger kids, grades means a sort of identity. That is the reason why I would push them as back as it possible, honestly. And there is also an extensive literature in education looking at how teachers should give grades because they are really, really, really bad for motivations. Really, really bad. I can give you a personal experience. I was very good in chemistry, till at some point I got a bad grade because my teacher discovered that I wasn't doing just my test, but I was doing also the test for two of my classmates. And instead of saying, okay, you are very good, you have been so good and so fast that you could finish your own test and making the test for other two classmates, she gave me a bad grade and never been in the class of chemistry ever again. And I think, as me, since I'm not special, as me, many other kids, many other students would feel exactly the same. Yeah, another kind of piece with the language models and the synthetic intelligence in the classroom. Yes, they're disembodied and there is that angle that we explored. Also, though, it's kind of like there's a new smart kid on the block. There's a new affordance, which is like, what year did this happen in? And that went from being something maybe that could be found in a book. More actions that have to be taken to walked library, open the book, etc. To maybe now search for it online to maybe now it's just going to be like, clap your hands and say it again, or maybe it's listening the whole time. And so how will the meta on developing personal agency and a feeling of being in the driver's seat in the education be cultivated? So people feel like they have tools to access and they have affordances at their disposal. Rather than feeling, I don't know how it could or would feel, but to be under the burden of much more knowledgeable, much faster agents that are directing the students. And teachers, I think, might also throw up their hands because they're not necessarily going to have recourse for technical questions either. And there will be this affordance at hand that will just bring out an essay or increasingly the audio and the visual. So how do we have a kind of fusion environments or different zones that help us wisely use that technology? I am a sci-fi fan. Okay, so I am not scared about anything of this guide. Bring in. I can deal with that. Honestly, after having been using for this point one year consistently, large language models, at the beginning I thought, okay, this is really a game changer. Right now I think, not so much at the end. The point is our approach. The point is how we teach people how to use them. And as for every other piece of technology, we all will use a different, in different way, in different circumstances. If I have to translate, I live in Germany, if I have to translate a piece that I'm writing for sending to, I don't know, an email to the mother of my daughter's friend, I can use JCPT and I will send a translation as it is. Okay. I won't pay too much attention. When I use JCPT for working instead, not only I check thousands of times the information that I have received because despite it might have faster access to the knowledge but doesn't have the right knowledge somehow. And this is something that you feel when it's right and when it's not. When you write a piece and you make JCPT, I make JCPT checking my grammar and my mistakes because I'm not mother tongue. When it gives me back, I always pay attention to put just no more than 10 lines because otherwise it restructures my sentence. And usually what comes out is nonsense, is changes completely meaning and stuff as such. So, yes, they are useful. We have to learn how to use them as we have learned to use to use. I remember in 2011. Yes, it was 2011 came out the paper. I think it's called the Google Effect. And they looked at how much we have in the opinion of the researchers, we have been losing memory because we started to use Google all the time instead of memorizing things, right? I see things in a different way. We haven't lost memory. We have learned how to categorize things that we want to remember in a different way. I don't have to store the information itself. I have to remember how to retrieve that information. And the same will be for JCPT and for the large language models. Also, sometimes ago, I was using JCPT and at some point, and this is the moment in which it was clear to me. I asked for some references about some toolmaking stuff for early on. And it provided me with a book of Kim Seralini, a book of Kim Seralini. And beyond, I know Kim perfectly. I know his work very, very, very nicely. And I looked at this book. This book doesn't exist. I'm 100% sure that this book. I think I looked at the title. I think this is the book that he should write, but this book doesn't exist. Because at the end of the day, I know better. Nice, yeah. Many funny things like that. And knowing where to find the information, whether the internal foraging or external niche modification, it's kind of always been the game. It's just a little bit modified now. Now, I'm sorry, I interrupt you. I'm now working on TXM on Extended Mind and Learning. And the active inference here has a very nice way to change the game of extra storage of information and so on. She's very interesting. Yes. I have a friend who says he has the largest collection of seashells in the world. He leaves them on all the beaches. Well, where do you go from here? Or what would you like to see happen? Or what would you suggest to people who are excited about pursuing education and active inference? I don't know, because as I said, I tried to create with judge EPT. I would show to you all the attempt to create an image of a classroom in which was applied the theory of cognition of active inference with the Montessori method. And I had lots of weird results because it's something that apparently nobody has tried yet. But what I would suggest is trying to make practical, to think in terms of design, to think in terms of affordances, to think in terms of everyday interactions, to think in terms of buildings, to think in terms of artifacts. This I think is going to be the next step. Awesome. Yeah, just to read a few comments, Joanne wrote, it brings the environment into the picture. We need to examine that more closely. Collaboration, not competition. And then JR wrote, it's like lifting memory from remembering the object to remembering the process. In a more complex context as the world evolves. Yes, yes. The good aspect, I think the point of putting together something like the Montessori method and active inference is understanding that learning is not about acquiring information, but about understanding the information, how to use the information. That is the game changer. Also, as my daughter says all the time, mama, today I reinvented mathematics. Amazing. Well, Laura, thank you very much for the work. It was definitely a topic that people, I feel like, wanted and want and need to explore how to be active in education and not get swamped in the tools and in all of these efforts and to kind of keep the hope alive for what education can be. Yes, yes. If we should put a little bit more ecology and sustainability inside, then we have the complete picture. Awesome. Well, till next time. Thank you very much. Thank you very much. Bye. Bye.