# Transcript — vanchurin-world-as-neural-network-2020 Source: https://www.youtube.com/watch?v=tAKaqojK0d8 Language: en Origin: Auto-generated (YouTube) --- we might be all just thinking about the same thing but from different angles and and eventually we can converge to something you know and it doesn't matter who gets there first what's important for us as a civilization is to get there to get the right answer [Music] [Music] so hello and welcome to a theory of everything i'm luis razzo the director of asem and host of this channel today i'm talking to vitali banturan a theoretical physicist and cosmologist at the university of minnesota who has just published a theory of everything based on the idea that the universe is a giant neural network of the many people that i've been privileged to have as guests professor rancheron ranks at the top in terms of the intellectual affinity of his work with the mission of this channel if the framework he's putting forth turns out to be an accurate description of nature it will almost certainly have important implications for everything that matters to humans i'd like to draw special attention to two things in particular first note how venturan's framework has directly testable practical implications for artificial intelligence and machine learning anybody working in these areas would do well to heed banjon's advice and work through his two papers on the topic you'll find a link to them in the video description second note how the symmetry manchurian is proposing between the information processing of neural networks and the physical and mathematical structure of nature is perfectly aligned with the conceptual symmetry we proposed in our inaugural lecture a link to which you'll find on the screen now there's a lot to be said i think about the fact that there's a growing number of theories of everything that are converging around a core set of basic ideas throughout his career yakir aharnav has been adamant and rightfully so in my opinion that in physics we should always strive to rise above the math to develop an intuitive picture of what's happening because only intuitive pictures will help us figure out the right questions to ask the mathematics to solve in this regard professor banteran has clearly risen above the math and he's done so in a way that in my opinion puts him in a unique position to ask and respond to the biggest and most important question of all which is what do we need to do to survive and thrive as a species i hope you're as intrigued as i am as always i placed a timestamp in the video description i've also added more details about professor vantaran's research interests and bio i hope you enjoy the exchange and thanks for listening professor banter and you're in minnesota and i'm in barcelona welcome to a theory of everything and thanks so much for doing this ah thank you glad to be on your show i'd like to start with one question uh with all of our guests it's not a trick question it's just um a plain and simple question do you agree or disagree that the worst thing that can possibly happen from a human perspective is for our species to go extinct tomorrow okay uh yeah i think it's an interesting question uh but by our species do you just refer to humans or all species on the planet earth human species from a human perspective yeah so i think it would be pretty bad so i agree it's maybe um it's a bad thing but it would probably be worse in my opinion if all species went extinct so somehow i feel attached to other species here on earth and i believe if we if we don't you know we're not worthy of uh you know uh advancing the the the intelligence of of the earth further of our solidization further but maybe there are other species that will take over and it wouldn't be that bad in my opinion so yeah it would be pretty bad but maybe the worst thing can happen then everybody goes extinct and that that's um that's my answer but i don't know if uh okay i answered your question i i may not attempt to it's a difficult question i think you know you have to think more about this really understand that yeah okay very good well thank you for the response i think yours is one of the first one that was extended most people say that it would be it would be the worst thing that could happen but i agree with you i think it would be even worse if everything died you know i think uh if we died but other things live maybe there would be a chance for more i mean yeah the evolution will take over again in my opinion right and then the other species will uh will develop you know um they might be smarter than us they'll probably you know they may survive for longer they may so i think we are in this together uh us humans and other species and we are just trying to you know um pass the great filter together we're trying to get to you know to the development that will allow us maybe to you know um go to mars as some people are trying to do or further explore the galaxy uh if i knew that humans would go extinct but other species will you know uh will win this race i i'm okay with that okay okay that's very interesting answer well thank you very much you've given me some some food for thought i think you're right as well okay well you recently published a very interesting paper on what you call cosmological uh neural networks before we get into the very interesting uh details of that can you tell us about your research and how you arrive to this and some of the people you've worked with etc yeah sure so uh originally i was interested in a very large-scale physics cosmology you know studying the universe and i got interested into that because my advisor my phd research advisor was alex lanking and he's he's an expert in cosmology and so the question that interested me the most at the time was probably question of uh inflation cosmic inflation uh later i i kind of understood that you cannot really talk about it without looking at them more a bigger picture which is um uh which has to do with eternal inflation or uh about multiverse pictures so so um that was uh very interesting for me but yet i have uncovered uh problems with this paradigm and uh you know i worked with andrew linda on this topic and then later with ellen guth with whom we uncovered some paradoxes that seem to be there are regardless of what you do and um so that was interesting learning experience for me and of course um i was fascinated by the results but i understood that unless we understand um the more the smaller scale physics better quantum gravity we may never be able to answer all those large-scale questions and so my interest is gradually shifted towards smaller scales trying to understand you know what is it how can we learn and model and understand um the quantum gravity in particular and there were very interesting ideas that circulating in the radical physics community um you know one very well-known idea the string theory later old people discovered the dualities and and one of the um scientists from whom i learned all this was annie saskatchewan and stanford so to name you another person who was very influential for um from my perspective and so gradually yeah i became interested in the small scale or you know uh how to reconcile quantum mechanics and gravity together and that that that was all and um but i kept encountering problems so uh and um so i guess uh my recent attempt to to resolve the problems was to uh step back and see whether we might actually um be wrong in trying to think that quantum mechanics is the right theory and gravity should come emerge from it uh maybe there are some you know deeper questions that needs to be answered even before we can talk about uh unification of the theories and so that brought me to the idea of um uh using neural networks uh to model physics and vice versa you know use uh physical tools to understand how neural networks work and so so yeah so to you know before going into details about this particular work you know to answer a question about what my interest research interests are that you know from the cosmology large scale sites it would be multiverse cosmic inflation dark matter dark energy or even more exotic objects like cosmic strings that are less popular now and on the you know smaller scale physics or higher energies i would be probably interested in duality mappings that people are discussing string theory uh framework uh and more recent ideas that are also very interesting about how to use quantum information theory to study um yeah but you know the recent step is said okay i mean maybe we need something else and then that's where i come with this new theory that you didn't give me permission to talk about yet but i don't know it's very very all of these topics are so fascinating well can you connect this then and how you arrive to the idea that the universe could be a giant neural network right so i think my first objective was not so ambitious all i wanted is to uh use toolbox that we have in physics uh to understand how deep neural networks work um uh how um and um basically you know usually deep neural network considered to be like black box it works but we don't really understand how things work and when they work uh there is a universal approximation theorem that gives us an idea why um the neural networks are so good in modeling but we don't really know when they would be efficient in uh doing them learning um uh and so so my idea was yes not so new is to take the toolbox of statistical physics and apply to study on neural networks the the main difference was probably that i took a bigger toolbox than others do and i uh you know partition functions and thermodynamics everything i know about physic i said what is it that we can use to model uh neural networks and and before i even started doing that i realized that that the whole um theory of neural networks has to be a little bit generalized so i generalized a little bit i allowed more freedom to something which is called a loss function or cost function i allowed more freedom for kind of architectures that you can have and uh once i did this that that was uh sufficient for me to start modeling it using statistical mechanics and then arriving in a thermodynamic limit so then you know it was able to formulate let's say second law of learning which is an analogous to the second law of thermodynamics first law of learning which is analogous to them personal thermodynamics and that was all great because you know it wasn't really physics but it was applying the physics tools uh to to model um you know something that we haven't computers you know something that runs on computers like artificial neural networks ask you a question analogously so are we talking about an analogous process are we talking about a physical process yeah so for now no so this first step there was no physical process there like you know you have neural networks they run on the computer but we there are still a bunch of physics tools um think about the math tools that we use in physics that can be useful for understanding um you know artificial neural networks machine learning and moreover you can not only understand how they work you can use the results to design better architecture to make artificial neural networks work better uh okay so but that was like one project but once i i finished it i saw that something interesting is arising there uh something i didn't quite expect i saw it in certain limits the artificial neural networks can be modeled by using quantum mechanics uh not the schrodinger equation that you hear about but modeling equations something that has to do with more of a hidden variable uh approach or vomiting mechanics approach to quantum mechanics so is it a coincidence or is is it is it maybe not or maybe there is something deep into that and so then then i went to this you know second question say okay so i kind of have my own understanding of how physics can be useful to understand machine learning can we reverse the water can we now use machine learning to model and and physical processes physical system and again i was still far away from you know the idea of a cosmological neural network it was just okay so we have physics we have known phenomena physics like quantum mechanics have general relativity uh can we model both of them maybe neural networks microscopic neural networks uh can can uh have emergent behavior which on one side will give you quantum mechanics on the other side will give you um general activity and gravity and maybe also it will give you observers you know um and and that would be great because um you know people often talk about unification of gravity and quantum mechanics but kind of uh put aside the problem of observers and it keeps coming back to us there is a measurement problem in quantum mechanics that has to do with observers there is a measure problem in cosmology which also has to do with observers so so i i always have this feeling that the three things might have to be unified together and and we just need a better mathematical framework for doing this and so that was the idea and then and i just uh took uh the mathematical model of um neural networks that i've developed and so can i apply it to now get those behaviors and yeah you know quantum mechanics was one limit that was there now closer to equilibrium so if you wait long enough that the system uh will behave and will be described its solution will be described by equation of quantum mechanics with hidden variables so there will be hidden variables that are you know if you don't know those hidden variables that is why your wave function can collapse in in different ways but if you knew them then you would know exactly how your experiment uh that would be the outcome of your experiment um yeah and and so so that was interesting and and the most interesting part for me was that i understood the meaning of the uh complex phase that you have uh in in the quantum mechanical state there is this phase that uh where does this quantum number uh or exponential to the um uh to the you know to the phase comes from it turns out it's related to thermodynamic properties and to free energy i'm not going to go into details but that was kind of was this was this the moment so we understand was this the moment when you said you know i'm onto something or at what point did you yeah yes that's the moment where you kind of uh you know i mean there's something interesting there there's i mean now i i kind of understand where quantum mechanics come from and i've a little bit bigger and understood where this uh quantum phase is coming from so yeah i mean maybe there was something different to that i still didn't have gravity at that time so it was just one of me and i didn't even have classical manhattan before you go into gravity can you tell us a little bit more about the hidden variables and yeah what role why is it why was it important in your mind yeah so so uh there are two types of variables that describe neural network things that like change moving parts and they're uh something that's called the state of neurons and the other one is is is the weights the connections the weights of the connections uh the weight matrix and then that and and so that two different variables and uh it was important to see that uh one of them can be modeled as quantum variables and so the weight matrix and the bias vector those are the ones that you can model as quantum variables now uh but what if you don't know about the other type of variables what if you don't know the the states of neurons but if you don't know what they are you know how would you model the evolution of your system uh if you do not have precise knowledge of the system and and and so and that's kind of that's why i connected i say well if you don't know them the best way for you to describe them is in terms of thermodynamics using the free energy concept okay but that free energy is uh due to the fact that we don't have a full knowledge of those hidden variables of the state of the small neurons individual neurons if we knew what they were evolution wouldn't be quantum it would be completely deterministic so yeah and that's where it's clicked and i kind of i thought well that's nice now i i have uh understanding microscopic understanding what those hidden variables might be but it is only in the limit where everything equilibrates so it's only in one limit that you see that behavior uh but perhaps for small system it doesn't take long time to equilibrate and that's why you know quantum mechanics is the right description of small systems where things have already equilibrate but that was just quantum mechanics we also know that classical mechanics uh you should have a classical description on larger scales so and and this framework uh the classical description is further away from the equilibrium so if you move away from the equilibrium then you have uh the classical equations of motion give you the correct description of the system and that's where i i thought that there's something really interesting there because what if we're moving even further away even further away from the economy what happens then uh but and so kind of classical mechanics and quantum was there but what about gravity so i you know it's always uh on the back of your head i mean gravity how do you connect any model you have how do you connect gravity to this because gravity seems to be very different so at that time um i think i borrowed some of the ideas of string theory uh important ideas of the string theory so in particular there is a bullet of action um that that underlies the evolution of uh and although that the strings are treated quantum mechanically there still the same action the same equations of motions emerge in my description so instead of having microscopic strings what i have is a microscopic chains of neurons so one neuron connected to another to the third so you see kind of the um the one-dimensional structures emerging them and their equations are similar to the ones that you have in string theory so what would be the difference between a one-dimensional string and a neuron yeah so there's a chain of neurons so neuron is just about yeah so the main difference was that the chain of neurons are uh have to be treated as a stochastic variables they're random variables they're they're like um you know they're not quantum ins in string theory the strings are quantum so here this the chains of neurons are not quantum at all but yet their approximate evolution is described by similar equations of motion so that that's a big difference but but the the reason why i have space-time emergent in my picture and and also general activity it's very similar to what you have in string theory so there are uh deep connections to string theory but their their you know calculations are are quite different do you borrow at all or do you yeah do you borrow it all from the ads cft correspondent yeah okay so so that comes next yeah so uh that's even um deeper idea that i had that i haven't quite finalized yet but the point there is that um you can model your system far away from the equilibrium that neural system and it is described well by gravity uh but then in the bulk inside of this system which model has gravity there are long chains of our neurons and so basically chains and neurons are not very efficient in doing calculations all they do you know they pass a signal from one to another and it goes from one boundary to another boundary so you have a very um sparse neural network where you don't have all neurons connected to one another so can you take this neural network and replace it uh with a dual neural network which is not sparse uh but which is described by the quantum mechanics and i think that is possible so basically you said dual neural network why dual right yeah dual meaning that you know in one case i some have some functional of some variables and and without going into details this function determines everything there is uh to to learn about the system all the observables okay but can i map it between other functional which has a completely different form yet when i studied the observables they are matched and so that's the yeah so mathematical structure is completely different right but yet when you you know try to extract the observables you can match this to that and that's kind of this type of duality is still uh emerging there but there it's a neural network duality in one case you have a very uh deep uh and very sparse neural network and another one you have uh very shallow and and densely connected neural network so uh so in a way that the you know holography not in the scft but holography uh where you are describing you know gravitational theory of those neural chains in the bulb is mapped to theory on the boundary where there are no long chains but you know everything basically connected earth that is present there uh now wow that's that's really really fascinating and deep that was interesting to observe so you kind of can manage that so yeah and that you know once you've done all these calculations and um even those calculations are not complete the one main uh challenge that i need to address is uh to see the emergence of symmetries uh more symmetries that i already encountered now physics of the 20th century was all about uh symmetries we've been discussing new and new symmetries and the standard model is full of symmetries there is a symmetry of uh the structural relativity general relativity so uh seeing how all those symmetries emerge is the biggest challenge so far i do see a symmetry of this quantum mechanical phase inversion there u1 i also see the symmetry of special relativity and generativity emerging but there has to be more and that kind of is like work in progress how can you get not just how you can get standard model how can you get uh dark matter how can you get dark you know all those things are very important but anyways once you have uh an idea that maybe it is possible to use neural networks to describe all different um you know maybe conflicting theories then the question is okay so maybe the whole thing isn't just a neural network and so that's what you would call a cosmological narrative everything is connected to everything you know some connections are strong some connections are weak uh and and okay so if that's the case uh so what can we say can we say that what you found is a kind of a cosmological symmetry or a overall general symmetry would that be an accurate um description yeah so once you just write down the neural network and you put all the random weights there's no symmetries so if there is a symmetry to emerge it has to take time through learning so i think you know uh saying that that there is this cosmological uh network doesn't imply that there is any symmetry but what what is a symmetry is when this cosmological network evolves towards you know a space time where the locality makes sense now you know uh you're further away from me so neural connection between me and you right now is weaker but you know things that uh next to me the neural connections between those degrees of freedom are stronger uh now it doesn't mean that you cannot have a strong connection of things far away okay you can but this is not a process that you would generically produce so this is something uh very out of equilibrium uh process that would give you strong connections between faraway places and maybe there are so maybe you know there is a strong connection between us because we are kind of discussing the same idea so we are kind of agreeing on something and so because of that you know that some microscopic neural uh neurons are now strongly connected because you know we've we've discussed something we've uh agreed on something uh and and but but once you talk about cosmological networks it's it's a bit stronger than that it's not just you know here people on earth but everything is connected right and some weights are strong some connections are strong some weak but it doesn't mean that the i cannot have a a strong weight connection to something very far away something that i haven't quite uh uh have the tools to experimentally identify but it doesn't mean they're not there uh and so and that kind of supports the idea of hidden variables people the main complaint of the hidden variables theories was that uh they have to be non-local um and but neural network is non-local you know every neuron can be connected to every neuron and their locality in this picture is just an emergent phenomena something that you have approximately but not not exact well that's what i mean that's an interesting idea the idea of emergence emergence if everything emerges then necessarily it would seem that everything is connected in some non-local way yeah is that yeah is that that's that's the idea right so everything but but some connections are weaker than others so uh you know you may not need to be connected to the degrees of freedom of the sun probably uh you are okay with with you know photons coming from there and you connect with them uh and you may not need to have any strong connections to uh you know molecules on andromeda or you know something far away uh and so but approximate uh notion of locality i think this is the right one okay but i think we shouldn't postulate it locality is something that in this picture is only approximate it's a very good degree and a good approximation uh but then but then okay you can of course take this further and and try to see what else can be emerging right so we've discussed something on very large scales you know gravity you know cosmological scales something on a very small scale quantum mechanics what about in between you know uh what about particles what about atoms uh you know biological cells observers can all those things be emergent through some kind of evolution now if you look at a microscopic picture of neural network you see that there are certain structures that survive longer and uh certain structures you know loops of neurons that get eliminated faster okay and so the more stable structures will survive longer and the less stable structures will get eliminated that's what i would call evolution right so i mean this this structures them in uh more feet to to to to stay to to leave and and you know we are a structure like that so when you ask me a question about you know humans and would they be sad if humans get extinct yeah i would be very sad about that but it doesn't mean that we don't have here on earth better structures that are more fit for survival and they deserve to survive further and maybe they we can just you know learn from us you know we'll get extinct they will survive and then they will win the race there's a realization very true very good point i think you're right i think in fact that reminds me of something that um juan maldena refers to as the principle of maximal life he calls the principle he calls the principle of least action he refers to it as the principle of maximal life and i find that very interesting especially coming from such a preeminent physicist yeah i think i think he says that on purpose and that's very good very good very good interesting point well i have another question i want to keep you talking about these things but if you can mix in the concept of a theory of everything is this a proposed theory of everything and first what is a theory of everything and using this as a proposed theory of everything so so i think it is just a buzz phrase right you know uh it's it's a you know but and i think everybody has uh their own opinion of what it actually means uh to which degree do we want it to be a theory of everything so for me it's a framework a framework that allows me to model things now the more things i can model with it the better it is now when i say everything well then in principle i should be able to model everything everything from the microscopic scale to the microscopic scale and if if there is a way for me to do that that then i can claim yeah it's a proposal for a theory of everything now if at some point i encounter that there's something that cannot be modeled or modeling with the tools that i have make things very very complicated and and uh you know accounts razer approach isn't really producing the desired result then then i say no maybe this is not the right way uh to to to model it uh but for the time being i do see um how this idea of uh neural network microscopic or cosmological network can model everything so i see that as a possibility uh and for the for that reason i think it's okay to call the theory of everything now of course once you do that uh many physicists will consider you a crackpot just because you use that phrase but i don't care i already have a tenure position and so on i just put whatever labels you want that's that's fantastic i i i applaud i applaud your work and your efforts fabulous i would say that in of the proposed theories of everything or so-called theories of everything i find this one to be the most interesting because assuming just assuming obviously we don't know yet but assuming that it on the right track if the universe is a cosmological uh neural network then it implies a symmetry between our own neural networks in our in our heads and the cosm and the cosmological or the the universe and if that symmetry is real then there's it's very nice because it implies a physical and mathematical connection that we can make and so yours would literally be a theory of everything because it would connect not only gr and quantum mechanics but also the things of the art and observers absolutely observers exactly observers and art and politics and i think it's fantastic sure sure sure yeah well i mean all of those phenomena art and politics and ethics whatever you all those are very complicated would be very complicated emergent phenomena so uh now of course we are now speculating right of course calculations was only really done for quantum mechanics and a little bit for general relativity but if we are to like uh open our minds and see what's possible in principle then yeah i think all those more complicated phenomena uh politics the ethics you know whatever should should be somehow modeled within this framework now can we do it efficiently so that we can make progress so that we can optimize our society so it functions better so we actually win the race our civilization wins the race and presumably there are other civilization was you know thinking about uh same issues but for now we don't know them uh and so it's just about how can we advance our knowledge in a more and more efficient way so i think that that kind of i think this is uh why i think maybe this theory of everything is more ambitious than others but it just came it was not intentional i wasn't saying about everything it was i wanted just to understand quantum mechanics and gravity and observers in some way well very fascinating well one of the reasons i asked the question to all of my guests the first question about human extinction is because from a human perspective obviously it's something we want to avoid so if we have 5 10 20 theories of everything obviously from a human perspective we want the one that is most promising in terms of avoiding human extinction right oh i agree yeah go ahead i think to me that makes imminent sense and in physics for example it's well known to attract uh the smartest people some of the smartest people in the world if not these smartest people and but there's no connection yet until now i think at least a slight connection between physics and those really big questions like how do we avoid killing ourselves it's it's well worth yeah i think physicists are embarrassed to ask deep questions uh we don't want uh people in you know uh behind our backs to make jokes about us you know that's why physicists don't like to talk to philosophers that these businesses don't like to talk to uh you name it i mean we're we want to physicists are comfortable within their own small group because they kind of exchange ideas using the same formulas and they are comfortable in this and and now you're asking how can you go you know come out of your comfort zone how can you start asking bigger questions uh it's hard it's hard and i understand why our physicist would have a hard time to linking uh yourself and start and start asking those questions uh but i think we have to do it if you i completely agree if we as a civilization are to win the race we gotta use everything i mean the more information you have uh the better theories you can produce and the idea for us is not to filter all of the other ideas that come to the physics community not to filter it until we are sure that those ideas are on the wrong track and so having many theories of everything i think it's a good thing okay the more the merrier uh but we should be able to communicate those ideas to each other and and i think maybe part of the problem is not to come up with an idea but to you know write an article and connect it to what was said before uh connected to string theory connected to quantum mechanics connected to things that we as businesses know and been trained with um so that's why if if a person who proposed the theory of everything can do that then then he he or she should do that now if that person who proposes a theory of everything just you know thinks that he does he uncovered the truth and that is it and and uh everybody else should just learn from him her uh that that's not productive uh you know communicating your ideas it's hard it's very hard you know you may have a clear picture in your head but being able to take it put it into the formulas and describe it to somebody else it's a challenge but that's the only way we can make progress as a society as a civilization well that's very very profound and very deep and i think that's one of the most important uh series of sentences that i've heard from uh for many physicists in in maybe ever i'm not sure that's very profound very nice thank you thank you very much for that now speaking of theories of everything um there are there are a growing number of theories of everything so why would why would you or how would you explain this what do you think this is the cause of this uh growth right so so i i thought about it actually and i think um um you should not be surprised that there is a growing number of theories of everything right now now why because for the last 40 years or so the theoretical physics was in the phase uh where we're proposing many ideas but those ideas were not yet confirmed by experiment so i call this phase as exploration phase we explore the ideas great ideas um you know string theory is one uh dualities is another great idea that uh the quantum computing quantum information so we were producing a lot of good ideas but they were not confirmed okay now that means that uh most likely we are coming to an end of the phase and something big is about to happen so if this was exploration exploration phase there'll be a production phase where we'll actually be producing theories that can make a better better prediction it can uncover the laws of physics better and better i think this is the next phase and so because we were in this explanation phase for so long many people feel that that the necessary tools to enter that next phase were not yet developed okay and and ui have a higher chance of succeeding if you work on it now and i think they're right so i think the people who are proposing theories and trying to think about um very wild ideas crazy is this is the time to do this right now uh and uh and that's why i think we have so many uh proposals unfortunately those different proposals do not receive as much attention um you know and they are not maybe not presented the right way maybe the scientific community is um you know conservative in accepting your ideas uh or a combination of two uh but but i think those are two reasons we come to the end of the exploration phase i'm pretty sure there are gonna be uh production phases where we're gonna encounter something big what is that we don't know now maybe it's neural physics i don't know maybe it's something else but we're getting there so that's my my view on it but again i might be wrong okay well thank you very much presumably you are uh leaning toward your own ideas about neural networks are there any other proposed theories of everything that are promising to you for example string theory or yeah so i i don't think i'm leaning towards my idea i just want to have one more idea there uh i'm i'm very happy would be very happy every string theory uh would be the right one why because we've invested so much time into this and energy i mean if that's the right one there is like uh you know there's a lot of work that's had been done it would be easier to take it into adopt it if we made this final big step which is still missing now there are other theories you know that are people proposing that haven't attracted as much attention now loop quantum gravity is another proportion again i know if that's the right framework okay i would be happy okay any framework would be happy but what i'm really hoping for that this step and our uh shift from the exploration phase to the production phase will happen soon now whether it's neural physics like i'm proposing or a string theory loop quantum gravity or you know some cellular automata model that would be great too you know there are proposals of uh how computers can run you know a little bit the computer's stimulation uh that wolfram proposed recently an idea of his theory of everything um hoofed worked on his interpretation of quantum mechanics based on the cellular automata i think i would be happy with anything at all i'm just a little bit sad that for so long for 40 years we've been in this exploration so i i think uh you know i'll be happy with anything but i'm i'm really hoping that this shift when we go to something where we actually apply the ideas and and model the phenomena that can be experimentally verified uh happens soon so i i would be sad that if we have to wait for another 40 years i mean we've been in this exploration every phase for too long so that's my very good that that's a very good point okay um we talked previously about the hidden variables and the interpretation of quantum mechanics of of boom the bohmian interpretation and that led me to connect this to the interpretation of who talks about the two times and we talk briefly about that can you can you explain a little bit about what you caught on about his interpretation of the two times and your initial impression right so so um i briefly i heard him giving a talk about that so uh i think he has a good understanding of what the challenges quantum mechanics are and uh how we should go about trying to address them uh now whether it's his proposal that's gonna kind of uh determine what the right way of looking quantic mechanics or or it's uh it's my proposal where i actually view it as an emerging phenomenon from neural network or maybe the two are connected or maybe there's something else that's connected to us that i don't know yet so that's kind of uh probably um is um a hard question to answer now without you know talking to him and discussing it in greater details and understanding uh what the proposals are so i don't think at this point i'm ready to give you the final verdict of what i think of this but i think it's very important that more and more people understand that there is something fundamentally wrong with our current view on quantum mechanics and it i hear it from different people you know who expressed us similar similar opinion uh weinberg as well and so it's very uh important to to see that the people start realizing uh that there's something that we haven't got right so the shut up and calculate approach to quantum mechanics worked great in the 20th century but maybe in the 21st century we should not just shut up but should start you know asking difficult questions and trying to answer them and maybe and that was part of the reason why i've taken this um measurement paradox in quantum mechanics and try to to use it as a motivation to model everything using neural networks um and and and then the final outcome of this that was bohmian approach is probably the right one at least in my opinion but um and not the many worlds for example that i liked before so uh and and i prefer now to think in terms of hidden variables okay and how has you i know your paper just barely was barely published but how has it been received so far and what are the next steps in your research right so next step is to publish it so that's been submitted it goes through the peer review hopefully i'll get some feedback soon uh now uh i haven't heard any any uh feedback which would be negative i received a lot of positive feedbacks and and from people from different branches on sizes and then they they're thinking that i'm on the right track um and stuff but of course you know i have regardless of how i'm maybe satisfied or dissatisfied how peer review works i have to go through this process and then i've heard people uh saying that well i don't trust peer review i'll just go my own way and do that's not productive okay so we have to follow even if you have a very crazy idea and even if it's you know why you know some people might not like it because after all you know as i said this is a very conservative they have their own ideas it's very hard to make them adjust to anything new and that's a good thing actually it's not a bad thing but if you are um you know if you can uh write a coherent paper express and explain and answer questions and give talks which is what i'm doing now is that then that's the way to go that's i think um so i i'm actually i think i probably received more feedback about this paper that i thought i hoped for and so and that's good and so i think i'm um i'm making progress excellent so it when after you get it published how do you assuming everything works out well and you get lots more feedback can you would you then make this into your central research project well of what uh you know how the research projects are determined right so i have like uh 10 questions that i currently want to answer and i kind of assign probabilities which one is more important or and which one is more likely to succeed uh right now uh their follow-up ideas of this work are certainly much more important than anything i've done in the past so there are really important questions that needs to be answered and and the uh good thing about this is that if i manage to answer those questions for example about symmetries that i mentioned that will help not only modeling physics and understanding how the universe works but also should help uh the machine learning people the people who actually design uh programs uh and you know train um robots to do their job so i think uh by viewing whatever you know by doing the calculation that can benefit both machine learning um and uh community and the physics community i think this kind of this is the reason why i think that that kind of projects are uh prioritized in my scale of important projects so yeah i'll definitely continue working on that okay well last question um can you what are the implications or can you recommend something for people who are working on artificial intelligence or neural networks what are the implications of your of your conjecture right so i think for people who are not physicists but uh looking at machine learning i would suggest them to go back to my previous paper that is called towards the theory of machine learning where i actually apply physics to model and understand how machine learning works uh and and if you if they are able to get through it which is it's a it's uh it's a difficult read i have to admit there is a lot of mathematics and they're very unusual people uh from a machine learning uh perspective but if they manage to do this at the end by the end uh they'll be a machinery that will allow them to design our neural networks design architectures that that should do better uh to to accomplish whatever learning tasks they're trying to do there is a um you know a function that you calculate and and that will tell you okay this a neural network will perform better use it to design your own uh code to to program so i think if if the machine learning community is able to accept that the amount of mathematics that i'm trying to throw in their direction that would benefit them dramatically now if they just want to do programming and then you know trying to guess what the right architecture should be then they can continue doing whatever they were uh doing but i think there is a lot of interesting math in there and a lot of interesting stuff and i think we can help each other now i am now collaborating with some of the people in the industry that have real data and so some of the um theories that i have now uh they're computer scientists of course but they have real data i can check and you're able to test you're able to test your right so that now i'm a theorist but now i have access to experimental data where i can test the theories and and first test the theory is to improve and solve whatever problems they have but then once i understand how the neural uh networks work i can now take it and apply it to physics i think this is kind of my idea from now on very fascinating very fascinating so you must be very excited at this point in your career i suppose i am i'm a happy person yes yes good for you i i said what i said last question but i have one more leonard suskin who you mentioned previously is also working on a lot of this a lot of these questions right and he mentioned quantum information theory and the potential of building a quantum like shell to test string theory can you connect your work or you mentioned previously quantum information theory right why is that interesting to your work yeah so um i was interested in quantum information independently uh i became more interested in quantum information because of language skin work where he tried to connect the complexity theory of quantum circuits to gravity that was very interesting from my idea now even more recently he tried to develop the thermodynamic notions for quantum circuits so the second law of uh complexity the the first law and and uh if you look at it closely you see that it's very much related to the thermodynamic concepts of learning that i was trying to develop uh now for one reason why it's uh it's related is because in the case of uh quantum circuits you can uh model it and think about this as a tensor networks where the things are connected in the network-like structure and the same thing you have uh for neural networks so so that the very graph-like structure in one case in that or network-like structure in both cases are similar but moreover lenny has also taken the approach of thinking about this in a thermodynamic limit thinking what kind of uh emergent behavior you would you would get so here's more interesting complexity i'm more interested in learning now but i do see that many formulas go inside and there are many uh things uh we might be just talking about this same thing but using different languages and then and then the two directions can benefit from one another so i don't consider this uh quantum information approach uh is unrelated i think it is related very interesting idea i think you're absolutely right in fact i think many of of what my gut feeling about the growth of the theories of everything is that many people are getting different pictures of perhaps things that are not completely unrelated wolfram's ideas uh it seems like they're they're or eric weinstein you're right but but we are too childish right now to understand how those ideas are connected and we might be all just thinking about the same thing but from different angles and and eventually we can converge to something you know and it doesn't matter who gets there first what's important for us as a civilization is to get there to get the right answer and to survive yes yeah i agree and to survive yeah we're in this race together we have to remember professor i want to respect your time and i'm so grateful for for this talk where can people or what do you recommend people to do to uh to follow you to follow your work etc yeah i mean if if you have uh they have general questions i'll be more than happy to answer by email um you know i'm also open to interviews sometimes if i'm not too busy but uh i think the best way is is to go through papers and uh and i think i've had many emails where people should try to go through papers and you know they get stuck on some equation and then uh if you're capable of doing this yeah start from uh you know this uh towards a theory of machine learning and then go to this second paper the the world as a neural network now if you but but the if that's if you are a scientist and but if you have more um you know philosophical questions i'm i'm happy to answer any questions you have that shouldn't be a problem fantastic i mean email is the best way to go okay fantastic well thanks so thanks again and have a great rest of the day thank you if you like what you heard stick around and let's reason together [Music]