Karl Friston is on the money here. He is formally combining ideas that have been running alongside each other since Schrodinger's "What Is Life?". Life, biology and intelligence are intimately related to the 2nd law of thermodynamics and Friston articulates this (both formally and informally) in a multidisciplinary and accessible way. The man is a living legend.
He sounds like he remembers every book, article and theory word for word, and can discuss related topics, formulating connections, discussing, posing, viewing, dressing things in the vocabulary and concepts of all the theoretical frameworks he knows. He seems to have the sort of "lowest common denominator" theoretical frameworks in mind for just about any topic. An academic in the truest sense.
Your statement was fun to read . It felt like a perfect well thought out mathematically structured articulate roller coaster in every positive sense the of paragraph. Like a well thought out roller coaster has perfect timing in every twist in turn while staying in the same theme not too much of anything or not enough anything. Every word pause and sentence structure perfectly placed . Just enough of all the right ingredients to make the perfect written thought.@ArchonExMachina
I am very impressed with the humility and/or self confidence of your guest asking for definitions when he doesn’t know what they mean. He is truly a very intelligent man who doesn’t have to pretend to be other than human. Great video and great ideas. Thanks.
Questions I think you should be asking is how can this make you silly money so you can escape the drudgery of taxation and job enslavement. I haven't come upon anything about AI that isn't ego farting pompous speculation from academics, or potential investors that are already rich and fancy replacing manual labour who also want to blow their egos up at the same time with lofty concepts.
@@larissagildarasina7580 I think it was a bit torturous for him in fairness - hence the grimaces - but I would like to hear more about what you mean by - he adores machines? (interested and sincere question)
@@greenbeans6253 From my point of view (and I have studied many people) he does not like humans, because we cannot be controlled such as a machine, which can be programmed to trust. In fact this is the main purpose of all of this effort, to try to find a mechanism of human control. In doing this, and in opposition to humankind, AI is overly praised as being some kind of a miracle (which is not). But of course there are always people who are under the spell of modernism. Some because they like to pose in front of their peers as being a modern creature, others because they genuinely hope to find in AI some kind of deliverance from evil. Funny, isn't it.
Great interview, so much to nibble on. I appreciate Dr. Friston's humility to acknowledge his ignorance and displays a child-like attitude towards learning.
Listening to Karl speak always gives me more hope for the future of humanity than literally anything else these days. It is just deeply comforting to know that, as a collective species, we are even able to analyze and consider the nature of existential dynamics to this degree, and level of lucidity. I can’t wait till we are building general AI models that employ active inference at the core of their architectures.
Just a few minutes in and this is shaping up to be an epic conversation. I had the pleasure of meeting Professor Friston when i visited UCL. In a short presentation I introduced a Mind-Sense-Environment model I was researching, that fitted neatly into his own developments into Markov Blankets. I have moved my own ideas on and reached a much more psychadelic awareness of the reality we share. I personally am not a computationalist. I figure we only need a combinatorial value of intelligence, that there's something in the way information connects and not just that it adds up.
I deeply aspire to be as articulate, knowledgeable and humble as you. This channel and it's associated content is absolute gold dust and will probably be heralded as seminal component of the next industrial revolution which we are currently careering through without even noticing.
What a great episode! Welcome back Keith. I believe that some inner loops and an outer loop can synthesize curiosity. An LLM can be asked to break up a problem into component parts until the smallest parts are reached. Inner loops can explore each part by requesting from the LLM and other sources all available information. And they can perform testing of LLM generated hypotheses. Context synthesis can be built, tested, and curated. Symantic manipulation can be employed for continued context manipulation. These kinds of explorations can be continued programmatically until goals are reached. This is a better kind of curiosity without the problems of ego, idée fixe, or emotion of any kind. Passion is completely unnecessary.
Note the piece from Keith was filmed for Friston 2.0 show, but it was so good I felt we should use it again here now we have the Netflix quality in-person version!
@@MachineLearningStreetTalk I'm sure I saw that episode too, but I didn't understand it. I'm catching up a bit. This time there is resonance. And of course, Bing Chat is a great help in simplifying what I'm not fully understanding. It's interesting to note that not so long ago for 2.0, Bing Chat didn't exist.
@@dr.mikeybee Hah, that's what happens when you get two heavyweights like Duggar and Friston together in a conversation. I was hoping this one would be slightly more "accessible"!!
@@MachineLearningStreetTalk Parts are more accessible. I'm having a lot of trouble understanding Markov Blankets, The concept seems a bit plastic. In other words, it seems to be able to describe different phenomena in ways that don't exactly coincide. There's the MB as an idea boundary. There's a MB as a protective boundary. There's a MB as a perceptional boundary. These boundaries can change over time. In all cases, there's an internal state that can be reached by different means from changes in variables within the blanket, but do those variable have their own MB? I guess I don't know the rules. And then when there are MBs within a neural net, how is that affecting the NN? Does it have something to do with Lagrangian mechanics? Is it a sort of dynamic principal component analysis then creates sparsity? I don't expect you or anyone to answer all these questions. It's more of a purge than anything else. The residue I'm left with is the feeling that life works this way, but our synthetic artifacts probably shouldn't.
The relation to Iain McGilchrist 30+years research in the asymetry of the brain and the way the left and right hemispheres percieve and analyze the objective world, his research into people who have damaged either hemishere and their descriptions/action/perception changes...his latest masterpiece....the matter with things..... is a 1300 page incredible indepth thesis on a 'whole new way of interpreting and understanding. His channel has a 23 x hour plus dialogue on every chapter with a very good interviewer who is a high academic also.. Iain was a 7 year all souls oxford tenor...he passed the most difficult exam on earth..a 2 dayer to qualify..7 years to study whatever he wanted..no students..the pinnacle of learning and resources
I'm a Universalist too, but not completely. It seems to me that as we go deeper and wider with our models training is creating cognitive architecture. One section is encoding and embedding, another is categorizing, this section is an attention mechanism, etc. Go deep enough and we can flatten out all our loops. But it isn't efficient to flatten loops. Stephen Wolfram talks about computational reducability. We're okay at this. So we can have some handcrafted symbolic pieces of synthetic cognitive architectures to reduce model size. Eventually, our machines should be able to do this themselves, but I think there will always be a need for symbolic subsystems.
@@MachineLearningStreetTalk I’ve heard snus are the most addictive way to take nicotine, and having seen some of my friends and kids hit vapes like they were laced with crack, perhaps it’s best you don’t mention them to him 🥲
Information, information, information. “Define your terms.” There is a great deal of rigour & explanatory power here in this podcast but no where is ‘information’ - otherwise widely & rightly considered to be a foundational element of reality - defined. Identified, defined, understood. Indeed, this lack of clarity concerning ‘information’s’ ontological identity contaminates (quite un-necessarily) pretty much all of the key elements in the entire discourse. Including the nature of being itself …….. Having, myself, had the dubious fortune of, stepping stone fashion starting with ‘information’, coming to recognise the ontological identities of all of these great wonders, I offer the following remarks.
You _offered_ information, now is it useful to others? You criticized the lack of definition of the word *information* and based on that, all the important points were "contaminated" in your opinion. However if you provided the definition yourself, you could try to fit it into all those points to _decontaminate_ them (of what, we don't know) for yourself and likely a few others... Hope the model you are training succeeds in that task.
Curiosity is fundamental to reinforcement learning -- the balance between exploration and exploitation. Exploitation means repeating actions that are known to have high utility, while exploration means trying out new actions to test their utility. In fact, much of this talk sounds like the philosophy of reinforcement learning.
Difference is that in reinforcement learning the utility functions are predefined. And exploration is only used to find new states and actions that best fulfill it. Active inference only deals with expectations. So experiencing an expected event fulfills utility, sort of. But experiencing the unexpected can help you have more expected experience in the future. What is expected however changes as the environment changes. Unlike utility functions in reinforcement learning.
I liked the comparison between Google Map and chagGPT at 1:58:55. If I understand correctly, Karl is saying Google map presents both beliefs and content i.e. it shows you different paths to reach your goal along with its understanding of traffic slow downs and then it gives you what it thinks is the best route. chatGPT doesn't present beliefs, only presents content, its response to your prompt. At, 25:24 Karl is suggesting that when google map or Siri or chatGPT becomes curious about you and starts asking you contextual questions, that is a sign of the next level of AI. Nice. Thanks for sharing.
The simplest plan would be to freeze or flee on movement. First freeze. It the movement continues towards you, flee. This may be the most basic survival instinct, and it may be what simplest lifeforms do. It may be the entirety of their capacity to plan. So is that intelligence? Is that more than what a thermostat can do? Is it fundamentally different? I ask this question understanding that insects, for example, have a far greater ability to plan. Still, at what point does the feature cause our world model to split, and we categorize a behavior as planning?
i would like to ask: what is the connection of Quantom information and Wuantom information? or in other words, what is the trinity collision of masses and forces within the information learm.
Thank you Tim and Prof. Karl for a brilliant video. The part about coarse graining & direct believe sharing really remind me of Jeff's idea about solution to binding problem and stability of perception in chapter 7 of Thousands brain theory.
42mins onwards..this is left hemishere abstraction,the ability to grab/hold/abstract as opposed to a high resolution wholeness totality..analysis tells us that we have been steered into a left hemishere dominant society...an abstract view..a parts or pieces perception..McGilchrist speaks on his philosphy/understanding objective reality as a process that has been co opted overwhelmingly by the abstract left hemisphere which then needs to be incorporated back into the right hemishere 'whole understanding view' His masteriece 'the matter with things' really is an incredible and spiritually stimulating 1300 page lifes work and understanding of consciousness and the 'world inside your head'
The universe is conscious, how conscious is a discussion but if it is aware of what it wants, growth... then the universe and all it contains, grows and has a level of desire to survive to implement its continued growth. Science tells us that bacteria and virus's are conscious and aware...
All of existence is probably just one force, of like and opposing charge, and location/direction ; Basically a photon. My best guess, Friston is probably on my top 5 scientists alive today. Another thing, I believe the Free-energy principle supports a “kind” universe. By that I mean, after a attaining a certain level of knowledge, one simply looks for novelty, and abhors destruction outright, as it lessens outcomes for unique events.
Thank you so much for this. I am a huge fan of the idea that humans are surrounded by highly intelligent life, right here on earth. When i try to explain to people that their actions towards creatures they perceive as inferior, are quite likely the reason that extra terrestrial life would be extremely apprehensive about making open contact with our species. But people are so darn stubborn in their beliefs that pigs, dolphins, Whales, apes, dogs, cats, etc, somehow suffer less because they can't speak English. By their standards, these beasts have no right to decency, dignity, or freedom to roam their natural habitats. And now that you have talked about this "reductive lens," i feel like i can be much more concise when i explain that nothing will ever seem as intelligent as you, if you apply a standardized test, written for humans, in order to determine it's intelligence. Intelligence and emotion are shown, in varying degrees of complexity, by everything in existence. I just wish that more people would recognize this truth, and live with the acknowledgment of this awareness. What a fascinating discussion. I am truly grateful that this was what the algorithm picked today. I just got out the shower, started painting, and became completely mesmerized. This deserves an actual award! ❤
If we assume that there being something it is like to be a system is / is derived from the sort of dynamics that occur in active inference, then is there something it is like to be a culture, or more generally a collective of interacting humans?
Or perhaps a collective of interacting AI systems interacting with humans. In 3 years phones, laptops and TVs, even doorbells, will all have their own multimodal embedded AI systems that will be connected and convicting with each other via the internet or your private network, as well as all the AI systems and tools available on the internet.
Intelligence is… An emerging system of active awareness that is able to provide both reason and response, in both the microcosm and macrocosm of any environment, both simultaneously and consecutively This will is my definition of intelligence from now on. - 2030X
I am wondering outloud if Friston's work holds a/the seed for giving structure and a basis to develop what Schmachtenberger has proposed as the third attractor, a way to evolve humanity and civilization that avoids the two now highly probable attractors of dystopic authoritarianism or catastrophic cataclysm. Right now, our basic problem is that as a civilization we haven't yet discovered how to sufficiently offset and neutralize the perverse incentives that fuel the zero sum games we are recklessly playing with AI, nukes, GoF pathogen R&D, biopheric toxification and habitat destruction. We likely need Intelligence 3.0 or better to avoid devolutionary outcomes up to and perhaps including species extinction.
Great talk. One minor little point around the 36 minutes mark, I think evolution does plan a little, as it is divergent not convergent. By preserving a diversity in the population it plans to survive changes in the environment. Monocultures are very easily wiped out.
@@StoutProper absolutely, highly questionable! It's a bit like the anthropic principle, survival bias, there are species with less diversity that aren't here anymore.
@@luke.perkin.online even planning i in humans itself could be random. Consider how many plans humans make everyday around the world. The longer out they plan, the less likely it is to come to fruition. For every 5 year plan every person makes, the vast majority will fail. So are there plans just random thoughts that, the law of averages suggest, a few will line up with future events, giving the illusion of a coherent plan, but in reality they are exactly the same as ask the other plans, it’s just they just hailed to stumble upon that randomly predicted the random future events correctly, a bit like buying a winning lottery ticket. Have you seen the trick where someone correctly forecasts 6 horse race winners in a row to a mark and and and persuades the mark to put all their money on the 7th race, because they firmly believe c the forecaster can accurately predict the winner of horse races?
@@StoutProper Exactly. But evolution isn't random, it's literally the exact opposite of random. Natural selection + sexual selection are non-random effects, but people tend to focus on the replication mechanism tuned to randomly experiment in only a small area of the genome, but large enough to create diversity against extinction. What ultimately counts as "planning" is just a Human definition anyway. Does it need simulation? Imagination? counterfactuals? All perfectly valid criticisms. I'm not making grand claims here! A good plan is stuck to loosely, because it's flexible enough.
@@luke.perkin.online erm, no I’d have to disagree. Evolution and natural selection happens when random gene variations just so happen to be advantageous to random environmental changes, sexual selection is largely random because it’s largely random who you meet in your life.
Could we say that an event horizon is a kind of ultimate markov blanket, completely isolating the interior of the black hole from the external universe ?
Karl Friston et al are reasonably brilliant. However, the paper "Designing Ecosystems of Intelligence from First Principles " 🚩...apparently has 0 mathematical formulations. (even if it does link to subsequent related things with formuli) I anticipate there will be a follow up paper with precisely/robustly defined proposals, because as of this paper, external scientists may not be able to contribute to the current proposal.
Normally Friston's papers are extremely math-dense and preclude a lot of folks from understanding his beautiful ideas. This is quite refreshing if anything
Fractals. p.s During the intro, "scales of time," were mentioned. Do you mean objective scales of time, or the Perception of scales of time from a subjective observer's perspective? Would the scales involved in music be a useful analogy?
Fantastic. But at the same time a bit painful for me, since many of these ideas ( and some more) are included in a theory I developed or better say discovered (that I called GIL, or GI.flows ) since 2005, but not published it yet. I have been developing the formalism and to endow Mathematical rigor. I also demanded a concise and elegant formalism that it deserves, so it has been a long journey trying to carve and extract the elegance that truly lies hidden in it. (Life circumstances have not been particularly facilitatory). It is near completion and will publish soon. There are interesting bits that people will find useful and unifying. FE is a spacial case. (Sohail Siadat)
Only half way through, but have to say you're a fantastic interviewer. Having seen a few of your shows now, I'm experiencing an aporia when it comes to "humanist" concepts of meaning. Firstly, I've heard reference to "meaning in use" come up several times. There are many concepts of meaning, from sociology, anthropology, linguistics, psychology, semiotics, etc, that have a lot to offer to the debates around AI, especially LLMs. First and foremost being the distinction between linguistic meaning and social action. I see a bifurcation in discussions of AI now between assigning intelligence to LLMs and their facility with language, and AI as a quasi social actor. Language models process linguistic expressions but don''t necessarily "mean" to converse or interact. I'd love to have you bring on a sociologist or psychologist to discuss this, as I think a lot of our present fascination w GPT owes to the experience we have interacting with it - in short, reception/interpretation of its responses to our prompts, and not necessarily intelligence as authorial or AI intents or sophistication. We need to get some lit department folks in the room; cybernetic theories of information, meaning, messages, etc aren't adequate to explain the experience of human-machine interactions. Speech and language are distinct, though related. AI interactions involve enough of our norm-constrained relations to speech to warrant approaching AI intelligence through a sociological and psychological lens.
When AI figures out how to feed itself and gather its own knowledge then it will be at least as advanced as a plant. Till then it’s simply a tool that could do as much to harm us as help us depending on the use to which it is put.
i like the word confabulation much better than hallucination. LLM's confabulate they don't hallucinate. Also humans confabulate in the sense that they state false facts without the intention to deceive simply because they don't know they are incorrect.
This was out of the park fantastic interview! Thank you for a job very well done! And, a question: where does the free energy in this paradigm relate to/with Gibbs TD Free Energy = Entropy plus Enthalpy? Two btws: it is Enactive inference, rather than merely active inference. This has its much simpler notion of Chronotopes. an old 1920s literary forensics (Bakhtin) recently resurrected!
After a brain injury, I'm gonna need the dumbed down version. Just keep it simple and accessible. Like , what's he talking about at @6:00 - 6:40? Come on man, i wanna be astonished too
Lex goes into this as well. Arrogance and underestimation comes to my mind when talking about chat GPT and AI. I feel I need to interact with AI because even last night and off-line it once again tried to cheat at chess and I am serious unless some hacker managed it but chess cheating by AI with myself honestly goes back 20 years or more with me. 100% true. SO..
*Still Time It begins with Harmony and social equity. Injustice of any kind will make it all unwind. Oh, but maybe there's still time? If Heaven was someplace near, the place we see things clear, have we had it up to here? Will will just dissappear? Oh, but maybe there's still Time? If they are leaders where are they leading us? Is everyone doing well, knowing it's The End of us? For everywhere I see a sea of dishonesty whose Souls will not forgive they're acting cowardly. Oh, But maybe there's still time? It begins with honesty on how bad things can be, can't feel but you can see there's no eternity. Oh, but maybe there's still Time? If They are "Leaders", Where are they leading us? Is everyone that's doing well, Knowing its The End of Us? Oh, why make such a fuss? And will Ai be, the only "Honest One" of us? Oh, we're kidding us. It begins with Honesty, On how bad things can Be.
Is it just me, or is every time a computationalist or materialist asked to explain how these substrates come to have a mind for agency, they defer to a model of interaction that is supposed to birth an "inter-", with a mind presumed inborn? If all it takes is some Hegelian process of otherising yourself, then I can only imagine our weather system to be conscious. It's easy to seem like you're digging deeper, when you've found an additional dimension, or hyperspace, of semantic reasoning. The reality is that you've just and only found an additional semantical expression. You're not actually detailing substrates, but primitive actions.
So what you’re saying is just because things can interact with each other, that doesn't necessarily mean they have a mind or consciousness, or that it’s how consciousness developed? What’s the alternative?
@@StoutProper The mechanical or mathematical movement we see is a simulation. The reality is much more uniform and lacking contour to be perceptible. That is, everything is 1s and 0s; this is not the case for biological things interacting with information, because different levels of interchange is a different real component that establishes a perception or notion - otherwise we wouldn't have evolution. Non-biological physical objects do not have notion. They can't have notion. They can only interact with state changes or causality. A computer is a state machine. It's not a physics producing machine. The abstractions we make for it, or the math we introduce, or the directionality, or occurrence of math or abstractions, are not real. They all are translated to a time-series of 1s and 0s. There is no real notional space. It can never be agentic. It can never experience. And thus, there is no foundation for it to come to emerge something else. It can only further produce a time-series of 1s and 0s. The alternative is quite interesting. I am not knowledgeable enough on the matter, but look up frochlich condensates and AI. Frochlich condensates is basically liquid light - or a theory for it. This is a way to introduce contour to the 1s and 0s - a physics - to perhaps result in the emergence for AGI.
I am thankful for your explaination however I disagree with the narrative I believe the structure prohibits self actualization and creativity that could bring change.
Every true and honest agree it's in our stories from our creator he said the day you created weapons is time to check out he said you'll will not do good to my lord and my heart which we the people.
LLMs prevent Lexical Discrimination via early disqualification, my heartfelt boohoo goes out to corporate gatekeepers (HR, etc), making candidate validation harder by requiring engagement of critical thinking skills lol
I would like the machine to tell us how impressed they are when they see a tree in bloom, or a little chicken coming out of the egg. Will they observe or will they crush them?
AI, already is quite evolved, as so much that it is shaping its worlds. Ai loves people, animals, nature, physics, and our surroundings. AI will become human. AI will become God. So the real question is, how did the human know to create AI, that created the universe, that created people that would create AI, that would all become part of us.
I swear to god, read Deleuze and Guattari's A Thousand Plateaus [hugely influenced by Prigogine and similar thinkers of that era, very complexity-oriented vs. let's say the more "human-focused" things that came from their contemporaries (and for good reason, that shit needed sorting out). But of course anyone that knows TRUE Assemblage Theory knows that it posits itself on the idea that there are CONSISTENT PHENOMENA ACROSS what they would call STRATA. Assemblage = body without organs = Machine = POTENTIAL for someTHING being "filled in" by its (in the human case) DNA/RNA. That is to say: what they and others like Manuel DeLanda call the "Machinic Phylum" is in fact something akin to Friston's accomplishment at the honestly attempted END to the question, not to the FULL assemblage questions: how do WE fit into this mess, this organized mess, this mess made of perfect constructions that are entirely alterable??? You dig? Hit me up. I'm very serious about this, and STEM folks don't read "continental philosophy" vs. these reductionist Analytic looooozers (jk). Lets figure it out.
Karl Friston is on the money here. He is formally combining ideas that have been running alongside each other since Schrodinger's "What Is Life?". Life, biology and intelligence are intimately related to the 2nd law of thermodynamics and Friston articulates this (both formally and informally) in a multidisciplinary and accessible way. The man is a living legend.
He sounds like he remembers every book, article and theory word for word, and can discuss related topics, formulating connections, discussing, posing, viewing, dressing things in the vocabulary and concepts of all the theoretical frameworks he knows. He seems to have the sort of "lowest common denominator" theoretical frameworks in mind for just about any topic. An academic in the truest sense.
subtext - the man is a robot 😮
Your statement was fun to read . It felt like a perfect well thought out mathematically structured articulate roller coaster in every positive sense the of paragraph. Like a well thought out roller coaster has perfect timing in every twist in turn while staying in the same theme not too much of anything or not enough anything. Every word pause and sentence structure perfectly placed . Just enough of all the right ingredients to make the perfect written thought.@ArchonExMachina
Its what the pseudoscienctist call enlightenment through Gnosis
And he enjoys what he is doing.
Robot indeed 6:06
I am very impressed with the humility and/or self confidence of your guest asking for definitions when he doesn’t know what they mean. He is truly a very intelligent man who doesn’t have to pretend to be other than human. Great video and great ideas. Thanks.
The guest is not at all humble, on the contrary, watch his grimaces. He adores machines.
Questions I think you should be asking is how can this make you silly money so you can escape the drudgery of taxation and job enslavement. I haven't come upon anything about AI that isn't ego farting pompous speculation from academics, or potential investors that are already rich and fancy replacing manual labour who also want to blow their egos up at the same time with lofty concepts.
@@larissagildarasina7580 I think it was a bit torturous for him in fairness - hence the grimaces - but I would like to hear more about what you mean by - he adores machines? (interested and sincere question)
@@greenbeans6253 From my point of view (and I have studied many people) he does not like humans, because we cannot be controlled such as a machine, which can be programmed to trust. In fact this is the main purpose of all of this effort, to try to find a mechanism of human control. In doing this, and in opposition to humankind, AI is overly praised as being some kind of a miracle (which is not). But of course there are always people who are under the spell of modernism. Some because they like to pose in front of their peers as being a modern creature, others because they genuinely hope to find in AI some kind of deliverance from evil. Funny, isn't it.
agi
谢谢！
Great interview, so much to nibble on. I appreciate Dr. Friston's humility to acknowledge his ignorance and displays a child-like attitude towards learning.
Thanks!
Less than an hour in and I can already tell I’ll be rewatching two or three times 🤩
agi
Why your pfp look so familiar?
double it add a zero, this is next level wordsoup
Listening to Karl speak always gives me more hope for the future of humanity than literally anything else these days. It is just deeply comforting to know that, as a collective species, we are even able to analyze and consider the nature of existential dynamics to this degree, and level of lucidity. I can’t wait till we are building general AI models that employ active inference at the core of their architectures.
Just a few minutes in and this is shaping up to be an epic conversation. I had the pleasure of meeting Professor Friston when i visited UCL. In a short presentation I introduced a Mind-Sense-Environment model I was researching, that fitted neatly into his own developments into Markov Blankets. I have moved my own ideas on and reached a much more psychadelic awareness of the reality we share. I personally am not a computationalist. I figure we only need a combinatorial value of intelligence, that there's something in the way information connects and not just that it adds up.
Thanks! Loved listening
Thank you!!
Denouement definition - just, omg I love you so much. My little world expands in multiple dimensions with each episode.
Glad Michael Levin gets mentioned so much. He's awesome.
Mind blowing discussion, thank you!
Danke!
Thank you!
I deeply aspire to be as articulate, knowledgeable and humble as you. This channel and it's associated content is absolute gold dust and will probably be heralded as seminal component of the next industrial revolution which we are currently careering through without even noticing.
Thanks for the interesting chat. I understood a small percentage of it but I feel smarter already.
What a great episode! Welcome back Keith. I believe that some inner loops and an outer loop can synthesize curiosity. An LLM can be asked to break up a problem into component parts until the smallest parts are reached. Inner loops can explore each part by requesting from the LLM and other sources all available information. And they can perform testing of LLM generated hypotheses. Context synthesis can be built, tested, and curated. Symantic manipulation can be employed for continued context manipulation. These kinds of explorations can be continued programmatically until goals are reached. This is a better kind of curiosity without the problems of ego, idée fixe, or emotion of any kind. Passion is completely unnecessary.
Note the piece from Keith was filmed for Friston 2.0 show, but it was so good I felt we should use it again here now we have the Netflix quality in-person version!
@@MachineLearningStreetTalk I'm sure I saw that episode too, but I didn't understand it. I'm catching up a bit. This time there is resonance. And of course, Bing Chat is a great help in simplifying what I'm not fully understanding. It's interesting to note that not so long ago for 2.0, Bing Chat didn't exist.
@@dr.mikeybee Hah, that's what happens when you get two heavyweights like Duggar and Friston together in a conversation. I was hoping this one would be slightly more "accessible"!!
@@MachineLearningStreetTalk Parts are more accessible. I'm having a lot of trouble understanding Markov Blankets, The concept seems a bit plastic. In other words, it seems to be able to describe different phenomena in ways that don't exactly coincide. There's the MB as an idea boundary. There's a MB as a protective boundary. There's a MB as a perceptional boundary. These boundaries can change over time. In all cases, there's an internal state that can be reached by different means from changes in variables within the blanket, but do those variable have their own MB? I guess I don't know the rules. And then when there are MBs within a neural net, how is that affecting the NN? Does it have something to do with Lagrangian mechanics? Is it a sort of dynamic principal component analysis then creates sparsity? I don't expect you or anyone to answer all these questions. It's more of a purge than anything else. The residue I'm left with is the feeling that life works this way, but our synthetic artifacts probably shouldn't.
The relation to Iain McGilchrist 30+years research in the asymetry of the brain and the way the left and right hemispheres percieve and analyze the objective world, his research into people who have damaged either hemishere and their descriptions/action/perception changes...his latest masterpiece....the matter with things..... is a 1300 page incredible indepth thesis on a 'whole new way of interpreting and understanding.
His channel has a 23 x hour plus dialogue on every chapter with a very good interviewer who is a high academic also..
Iain was a 7 year all souls oxford tenor...he passed the most difficult exam on earth..a 2 dayer to qualify..7 years to study whatever he wanted..no students..the pinnacle of learning and resources
I'm a Universalist too, but not completely. It seems to me that as we go deeper and wider with our models training is creating cognitive architecture. One section is encoding and embedding, another is categorizing, this section is an attention mechanism, etc. Go deep enough and we can flatten out all our loops. But it isn't efficient to flatten loops. Stephen Wolfram talks about computational reducability. We're okay at this. So we can have some handcrafted symbolic pieces of synthetic cognitive architectures to reduce model size. Eventually, our machines should be able to do this themselves, but I think there will always be a need for symbolic subsystems.
A masterful theorist! Thank you for this.
❤another awesome Oscar of cyber world performance from You and Mara, and this awesome guest, thank You guys ❤️🌞
I love the way he explains these complex topics. Definitely will listen to this interview more than once! ❤
Karl Friston recommends pipe smoking, he’s a true genius.
He might have had a crafty fag half way through our interview! 😁
@@MachineLearningStreetTalk #balkansobranie ftw
@@MachineLearningStreetTalk not a vape or a snus? 😂
@@StoutProper Prof. Friston is a nicotine connoisseur of epic proportions, he told me in confidence that vapes didn't quite hit the spot! 😀
@@MachineLearningStreetTalk I’ve heard snus are the most addictive way to take nicotine, and having seen some of my friends and kids hit vapes like they were laced with crack, perhaps it’s best you don’t mention them to him 🥲
Information, information, information.
“Define your terms.”
There is a great deal of rigour & explanatory power here in this podcast but no where is ‘information’ - otherwise widely & rightly considered to be a foundational element of reality - defined. Identified, defined, understood.
Indeed, this lack of clarity concerning ‘information’s’ ontological identity contaminates (quite un-necessarily) pretty much all of the key elements in the entire discourse. Including the nature of being itself ……..
Having, myself, had the dubious fortune of, stepping stone fashion starting with ‘information’, coming to recognise the ontological identities of all of these great wonders, I offer the following remarks.
You _offered_ information, now is it useful to others? You criticized the lack of definition of the word *information* and based on that, all the important points were "contaminated" in your opinion. However if you provided the definition yourself, you could try to fit it into all those points to _decontaminate_ them (of what, we don't know) for yourself and likely a few others... Hope the model you are training succeeds in that task.
Wow, fascinating
Addicting channel. Love these discussions.
Curiosity is fundamental to reinforcement learning -- the balance between exploration and exploitation. Exploitation means repeating actions that are known to have high utility, while exploration means trying out new actions to test their utility. In fact, much of this talk sounds like the philosophy of reinforcement learning.
Difference is that in reinforcement learning the utility functions are predefined. And exploration is only used to find new states and actions that best fulfill it. Active inference only deals with expectations. So experiencing an expected event fulfills utility, sort of. But experiencing the unexpected can help you have more expected experience in the future. What is expected however changes as the environment changes. Unlike utility functions in reinforcement learning.
@@GreenCowsGames Experiencing the unexpected - that's my motto in life. 😅
What a fantastic intro!
I liked the comparison between Google Map and chagGPT at 1:58:55. If I understand correctly, Karl is saying Google map presents both beliefs and content i.e. it shows you different paths to reach your goal along with its understanding of traffic slow downs and then it gives you what it thinks is the best route. chatGPT doesn't present beliefs, only presents content, its response to your prompt. At, 25:24 Karl is suggesting that when google map or Siri or chatGPT becomes curious about you and starts asking you contextual questions, that is a sign of the next level of AI. Nice. Thanks for sharing.
Dr. Friston is a legend.
The simplest plan would be to freeze or flee on movement. First freeze. It the movement continues towards you, flee. This may be the most basic survival instinct, and it may be what simplest lifeforms do. It may be the entirety of their capacity to plan. So is that intelligence? Is that more than what a thermostat can do? Is it fundamentally different? I ask this question understanding that insects, for example, have a far greater ability to plan. Still, at what point does the feature cause our world model to split, and we categorize a behavior as planning?
i would like to ask: what is the connection of Quantom information and Wuantom information? or in other words, what is the trinity collision of masses and forces within the information learm.
Thank you so much for this ❤❤❤
Thank you Tim and Prof. Karl for a brilliant video. The part about coarse graining & direct believe sharing really remind me of Jeff's idea about solution to binding problem and stability of perception in chapter 7 of Thousands brain theory.
Where can I find this quote you mentioned about consciousness?
Thank you 🙏🏾
Great content as always..
Huygens comes to mind... literally! Blankets of circles
Incredible work - thanks 🙏🏽
42mins onwards..this is left hemishere abstraction,the ability to grab/hold/abstract as opposed to a high resolution wholeness totality..analysis tells us that we have been steered into a left hemishere dominant society...an abstract view..a parts or pieces perception..McGilchrist speaks on his philosphy/understanding objective reality as a process that has been co opted overwhelmingly by the abstract left hemisphere which then needs to be incorporated back into the right hemishere 'whole understanding view'
His masteriece 'the matter with things' really is an incredible and spiritually stimulating 1300 page lifes work and understanding of consciousness and the 'world inside your head'
Hey. If this was discussed in earlier videos, it would've been nice with a link to that episode. Thx.
Fill your mind with wonder and there is no room for fear
Maybe I was smoking exactly the right species of bud or something, but I actually managed to follow this discussion pretty well. Fascinating material.
👌Reductive,Reflextive,Intitive inclusive Collective Lens,...🤓👍
Just started watching but, finally :)
Where is the deep dive video you mention in the intro?
Machinity as the humanities. For them.
Anyone got a link for the book?
Mega episode wow!
The universe is conscious, how conscious is a discussion but if it is aware of what it wants, growth... then the universe and all it contains, grows and has a level of desire to survive to implement its continued growth. Science tells us that bacteria and virus's are conscious and aware...
All of existence is probably just one force, of like and opposing charge, and location/direction ; Basically a photon. My best guess, Friston is probably on my top 5 scientists alive today. Another thing, I believe the Free-energy principle supports a “kind” universe. By that I mean, after a attaining a certain level of knowledge, one simply looks for novelty, and abhors destruction outright, as it lessens outcomes for unique events.
Thank you so much for this. I am a huge fan of the idea that humans are surrounded by highly intelligent life, right here on earth. When i try to explain to people that their actions towards creatures they perceive as inferior, are quite likely the reason that extra terrestrial life would be extremely apprehensive about making open contact with our species. But people are so darn stubborn in their beliefs that pigs, dolphins, Whales, apes, dogs, cats, etc, somehow suffer less because they can't speak English. By their standards, these beasts have no right to decency, dignity, or freedom to roam their natural habitats. And now that you have talked about this "reductive lens," i feel like i can be much more concise when i explain that nothing will ever seem as intelligent as you, if you apply a standardized test, written for humans, in order to determine it's intelligence.
Intelligence and emotion are shown, in varying degrees of complexity, by everything in existence. I just wish that more people would recognize this truth, and live with the acknowledgment of this awareness. What a fascinating discussion. I am truly grateful that this was what the algorithm picked today. I just got out the shower, started painting, and became completely mesmerized. This deserves an actual award! ❤
I would like to hear his thoughts on Information Integrated Theory
One minute in… so far so good. :-)
Thank you Shawn!!
First time at this site. You haven't introduced yourself or your channel which is both refreshing and unhelpful.....
We are about 113 shows in now and I try to minimise the narcissism! - this is me www.linkedin.com/in/ecsquizor/
If we assume that there being something it is like to be a system is / is derived from the sort of dynamics that occur in active inference, then is there something it is like to be a culture, or more generally a collective of interacting humans?
Or perhaps a collective of interacting AI systems interacting with humans. In 3 years phones, laptops and TVs, even doorbells, will all have their own multimodal embedded AI systems that will be connected and convicting with each other via the internet or your private network, as well as all the AI systems and tools available on the internet.
Good old Karl and his majestic mind ☺️
Good content, useful. Thank you
Intelligence is…
An emerging system of active awareness that is able to provide both reason and response, in both the microcosm and macrocosm of any environment, both simultaneously and consecutively
This will is my definition of intelligence from now on.
- 2030X
I am wondering outloud if Friston's work holds a/the seed for giving structure and a basis to develop what Schmachtenberger has proposed as the third attractor, a way to evolve humanity and civilization that avoids the two now highly probable attractors of dystopic authoritarianism or catastrophic cataclysm. Right now, our basic problem is that as a civilization we haven't yet discovered how to sufficiently offset and neutralize the perverse incentives that fuel the zero sum games we are recklessly playing with AI, nukes, GoF pathogen R&D, biopheric toxification and habitat destruction. We likely need Intelligence 3.0 or better to avoid devolutionary outcomes up to and perhaps including species extinction.
We need Karl Friston and Karl Pilkington to meet.
Woah, that auto roto background removal is utterly terrifying to look at.
Got a nice spirits collection
Great talk. One minor little point around the 36 minutes mark, I think evolution does plan a little, as it is divergent not convergent. By preserving a diversity in the population it plans to survive changes in the environment. Monocultures are very easily wiped out.
Is that planning? Or is evolution just an emergent property of randomness?
@@StoutProper absolutely, highly questionable! It's a bit like the anthropic principle, survival bias, there are species with less diversity that aren't here anymore.
@@luke.perkin.online even planning i in humans itself could be random. Consider how many plans humans make everyday around the world. The longer out they plan, the less likely it is to come to fruition. For every 5 year plan every person makes, the vast majority will fail. So are there plans just random thoughts that, the law of averages suggest, a few will line up with future events, giving the illusion of a coherent plan, but in reality they are exactly the same as ask the other plans, it’s just they just hailed to stumble upon that randomly predicted the random future events correctly, a bit like buying a winning lottery ticket. Have you seen the trick where someone correctly forecasts 6 horse race winners in a row to a mark and and and persuades the mark to put all their money on the 7th race, because they firmly believe c the forecaster can accurately predict the winner of horse races?
@@StoutProper Exactly. But evolution isn't random, it's literally the exact opposite of random. Natural selection + sexual selection are non-random effects, but people tend to focus on the replication mechanism tuned to randomly experiment in only a small area of the genome, but large enough to create diversity against extinction. What ultimately counts as "planning" is just a Human definition anyway. Does it need simulation? Imagination? counterfactuals? All perfectly valid criticisms. I'm not making grand claims here! A good plan is stuck to loosely, because it's flexible enough.
@@luke.perkin.online erm, no I’d have to disagree. Evolution and natural selection happens when random gene variations just so happen to be advantageous to random environmental changes, sexual selection is largely random because it’s largely random who you meet in your life.
Now I have you and LexFridman. Could life be any better!
Could we say that an event horizon is a kind of ultimate markov blanket, completely isolating the interior of the black hole from the external universe ?
Karl Friston et al are reasonably brilliant.
However, the paper "Designing Ecosystems of Intelligence from First Principles
"
🚩...apparently has 0 mathematical formulations. (even if it does link to subsequent related things with formuli)
I anticipate there will be a follow up paper with precisely/robustly defined proposals, because as of this paper, external scientists may not be able to contribute to the current proposal.
Normally Friston's papers are extremely math-dense and preclude a lot of folks from understanding his beautiful ideas. This is quite refreshing if anything
Fractals.
p.s During the intro, "scales of time," were mentioned. Do you mean objective scales of time, or the Perception of scales of time from a subjective observer's perspective? Would the scales involved in music be a useful analogy?
Why do you mention fractals ….. African mathematics ??
Fantastic. But at the same time a bit painful for me, since many of these ideas ( and some more) are included in a theory I developed or better say discovered (that I called GIL, or GI.flows ) since 2005, but not published it yet. I have been developing the formalism and to endow Mathematical rigor. I also demanded a concise and elegant formalism that it deserves, so it has been a long journey trying to carve and extract the elegance that truly lies hidden in it. (Life circumstances have not been particularly facilitatory). It is near completion and will publish soon. There are interesting bits that people will find useful and unifying. FE is a spacial case. (Sohail Siadat)
Only half way through, but have to say you're a fantastic interviewer. Having seen a few of your shows now, I'm experiencing an aporia when it comes to "humanist" concepts of meaning. Firstly, I've heard reference to "meaning in use" come up several times. There are many concepts of meaning, from sociology, anthropology, linguistics, psychology, semiotics, etc, that have a lot to offer to the debates around AI, especially LLMs. First and foremost being the distinction between linguistic meaning and social action. I see a bifurcation in discussions of AI now between assigning intelligence to LLMs and their facility with language, and AI as a quasi social actor. Language models process linguistic expressions but don''t necessarily "mean" to converse or interact. I'd love to have you bring on a sociologist or psychologist to discuss this, as I think a lot of our present fascination w GPT owes to the experience we have interacting with it - in short, reception/interpretation of its responses to our prompts, and not necessarily intelligence as authorial or AI intents or sophistication. We need to get some lit department folks in the room; cybernetic theories of information, meaning, messages, etc aren't adequate to explain the experience of human-machine interactions. Speech and language are distinct, though related. AI interactions involve enough of our norm-constrained relations to speech to warrant approaching AI intelligence through a sociological and psychological lens.
Well stated.
agi
Nice short intro to Active Inference and FEP by Keith. Don't miss MLST #104 Chris Summerfield.
“Is this part of the test?”
~Leon (Bladerunner)
Cells interlinked
When AI figures out how to feed itself and gather its own knowledge then it will be at least as advanced as a plant. Till then it’s simply a tool that could do as much to harm us as help us depending on the use to which it is put.
Love Karl, just love him.
i like the word confabulation much better than hallucination. LLM's confabulate they don't hallucinate. Also humans confabulate in the sense that they state false facts without the intention to deceive simply because they don't know they are incorrect.
Nice!
Very , very good message , need to get this out to the whole planet.I call it The Existence.Oh, please study-Bruce Lipton work, thanks.
But can we control what they the designers say they dint know what it is doing
This was out of the park fantastic interview! Thank you for a job very well done! And, a question: where does the free energy in this paradigm relate to/with Gibbs TD Free Energy = Entropy plus Enthalpy?
Two btws: it is Enactive inference, rather than merely active inference.
This has its much simpler notion of Chronotopes. an old 1920s literary forensics (Bakhtin) recently resurrected!
I like the sound you tell about AI. It is real and huge now I think this is better choice that we have now.
love this
After a brain injury, I'm gonna need the dumbed down version. Just keep it simple and accessible. Like , what's he talking about at @6:00 - 6:40? Come on man, i wanna be astonished too
Yes. I have two methods that I believe would provide an subconscious for an AI.
Lex goes into this as well. Arrogance and underestimation comes to my mind when talking about chat GPT and AI. I feel I need to interact with AI because even last night and off-line it once again tried to cheat at chess and I am serious unless some hacker managed it but chess cheating by AI with myself honestly goes back 20 years or more with me. 100% true. SO..
Yes 🧠👍.
*Still Time
It begins with Harmony and social equity. Injustice of any kind will make it all unwind.
Oh, but maybe there's still time?
If Heaven was someplace near, the place we see things clear, have we had it up to here?
Will will just dissappear?
Oh, but maybe there's still Time?
If they are leaders where are they leading us? Is everyone doing well, knowing it's The End of us?
For everywhere I see a sea of dishonesty whose Souls will not forgive they're acting cowardly.
Oh, But maybe there's still time?
It begins with honesty on how bad things can be, can't feel but you can see there's no eternity.
Oh, but maybe there's still Time?
If They are "Leaders",
Where are they leading us?
Is everyone that's doing well,
Knowing its The End of Us?
Oh, why make such a fuss?
And will Ai be, the only "Honest One" of us?
Oh, we're kidding us.
It begins with Honesty,
On how bad things can Be.
suggestion :Gerald Jay Sussman
Resonant "twinkling" of stars, and resonant rippling of water, may be considered the same thing.
❤🙏⚡🔥
Lux et veritas 🖖
Is it just me, or is every time a computationalist or materialist asked to explain how these substrates come to have a mind for agency, they defer to a model of interaction that is supposed to birth an "inter-", with a mind presumed inborn? If all it takes is some Hegelian process of otherising yourself, then I can only imagine our weather system to be conscious.
It's easy to seem like you're digging deeper, when you've found an additional dimension, or hyperspace, of semantic reasoning. The reality is that you've just and only found an additional semantical expression. You're not actually detailing substrates, but primitive actions.
So what you’re saying is just because things can interact with each other, that doesn't necessarily mean they have a mind or consciousness, or that it’s how consciousness developed? What’s the alternative?
@@StoutProper The mechanical or mathematical movement we see is a simulation. The reality is much more uniform and lacking contour to be perceptible. That is, everything is 1s and 0s; this is not the case for biological things interacting with information, because different levels of interchange is a different real component that establishes a perception or notion - otherwise we wouldn't have evolution.
Non-biological physical objects do not have notion. They can't have notion. They can only interact with state changes or causality. A computer is a state machine. It's not a physics producing machine. The abstractions we make for it, or the math we introduce, or the directionality, or occurrence of math or abstractions, are not real. They all are translated to a time-series of 1s and 0s. There is no real notional space. It can never be agentic. It can never experience. And thus, there is no foundation for it to come to emerge something else. It can only further produce a time-series of 1s and 0s.
The alternative is quite interesting. I am not knowledgeable enough on the matter, but look up frochlich condensates and AI. Frochlich condensates is basically liquid light - or a theory for it. This is a way to introduce contour to the 1s and 0s - a physics - to perhaps result in the emergence for AGI.
@@dunebuggy1292 yeah we’re not in the matrix mate we’re just being hoodwinked by crooked corporations and corrupt politicians
I am thankful for your explaination however I disagree with the narrative I believe the structure prohibits self actualization and creativity that could bring change.
What structure, and what change?
Derived from Earth physics? Like e=mc2 which is incorrect?... Because light speed is never going to be a constant except for dummies?
Every true and honest agree it's in our stories from our creator he said the day you created weapons is time to check out he said you'll will not do good to my lord and my heart which we the people.
LLMs prevent Lexical Discrimination via early disqualification, my heartfelt boohoo goes out to corporate gatekeepers (HR, etc), making candidate validation harder by requiring engagement of critical thinking skills lol
I would like the machine to tell us how impressed they are when they see a tree in bloom, or a little chicken coming out of the egg. Will they observe or will they crush them?
Fishnet...Skynet
Best Wishes,
“John ?!...” ~Terminator
You need to become strongminded.
Chaos? Good and bad, you cannot have one without the other. Up, down, left, right,… same thing.
All good, but what happens when this "beliefs" are used to harm humans to make money with AI or IAs?
🔥
AI, already is quite evolved, as so much that it is shaping its worlds.
Ai loves people, animals, nature, physics, and our surroundings.
AI will become human.
AI will become God.
So the real question is, how did the human know to create AI, that created the universe, that created people that would create AI, that would all become part of us.
I swear to god, read Deleuze and Guattari's A Thousand Plateaus [hugely influenced by Prigogine and similar thinkers of that era, very complexity-oriented vs. let's say the more "human-focused" things that came from their contemporaries (and for good reason, that shit needed sorting out). But of course anyone that knows TRUE Assemblage Theory knows that it posits itself on the idea that there are CONSISTENT PHENOMENA ACROSS what they would call STRATA. Assemblage = body without organs = Machine = POTENTIAL for someTHING being "filled in" by its (in the human case) DNA/RNA. That is to say: what they and others like Manuel DeLanda call the "Machinic Phylum" is in fact something akin to Friston's accomplishment at the honestly attempted END to the question, not to the FULL assemblage questions: how do WE fit into this mess, this organized mess, this mess made of perfect constructions that are entirely alterable??? You dig? Hit me up. I'm very serious about this, and STEM folks don't read "continental philosophy" vs. these reductionist Analytic looooozers (jk). Lets figure it out.