hello everyone welcome to another episode of Q&A about his the future of Science and Technology I am still not in my natural habitat so uh technical things could go wrong here all right let's see all kinds of questions that we have uh saved up here one from pavl here can AI be swallowed by more advanced AI by feeding it virtual input the motivation could be increased efficiency of larger AIS well let's see I mean the um the question is computation can do all kinds of things the issue is to align computation with the kinds of things that we humans care about and want to do things with and the main sort of idea in modern times is sort of train the AIS with lots of material made by humans and you know there's maybe I don't know 5 billion maybe 10 billion web pages that are sort of reasonably human written May maybe if you include all of the sort of hidden stuff there's maybe a 100 billion Pages worth of of things humans have written there's there's zillions of videos things like this there's lots of material that kind of shows what we humans care about and kind of Paints the picture of the kind of human relevant world and that's a that's a place where if you want AIS to do kind of human relevant things then you've got to kind of feed them human relevant training material that they will then imitate so to speak if you want them to do things that are related to for example the way the physical world works then having them observe the physical world and say oh I've I've grown up you know learning about how chemical reactions work or how proteins fold or something like this and that's my experience of things there's a question it's not completely obvious to what extent the kind of the idea of just train with what's already known is enough to tell you what will come next so to speak with with a lot of the things in the kind of human world what we what we're finding is that we humans possibly because our brains are not that different from neuron Nets that are sort of put into AIS that the things that that kind of we care about and that we are able to do are things that can readily be reproduced by kind of a trained uh you know a trained neuronet kind kind of approach whether that's true of the things that exist in the physical world is much less clear it's a little bit complicated because there are things that are intrinsic to the physical world where the physical world is doing what it does and there are aspects of the physical world that we humans have decided we care about like for example if we look at I don't know clouds in the sky the detailed shape of clouds is something that we humans don't usually make much comment on other than to say you know that one looks like a chicken something but we're we're um uh that's something but on the other hand the question of you know is it raining right now is something that we humans notice and so it could be the case that even though the underlying physics that's making the Billings of the clouds and so on is not something that is readily aable so to speak that is readily sort of extrapolatable from from training data that it's still the case that something like will it rain or not is and that's the part we humans care about so this question about whether sort of things in the in the physical world are what one might call sort of aable um is is it could be the case that at some intrinsic level they're not in fact that's my belief is that they're not for all sorts of reasons but that it's possible that the ones of human relevance are more likely to be I'm not clear whether they are or aren't but that's not something to to to ask about sort of the underlying processes we're running into this phenomenon of computational irreducibility that I seem to talk about a lot and that I kind of came up with in the 1980s um this idea that just because you know the rules for a system doesn't mean you can immediately kind of jump ahead and say everything about what the system does you are in general have to just go through sort of the irreducible computational process of seeing each step of what happens and if you're an AI that that's sort of really trying to just say is this something a bit like what I've seen before oh then I know what's going to happen that's not good enough if if you have computational irreducibility computational irreducibility you have to follow the steps and there will always be surprises and those surprises will be things you can't deduce from just saying well what have I seen before let me train on what I've seen before you know as a very practical matter I was just looking actually at um some the ability to predict like computational processes like cellular autometer with things like neuron Nets Transformers all the fancy technology and so on and the answer is it does kind of sort of okay in some cases and when things get complicated and when sort of unexpected things start to happen in these computational processes the kind of the AI predictors just sort of fall over and don't work now you know there are all of questions about what it's going to take to uh kind of get get to um uh to different levels of kind of sort of humanlike performance I mean that's a thing often of interest to us humans is can we get an AI to do humanlike things again separate from can we have the AI predict the positions of planets or something at some point in the future that is driven by physics the can we get an AI to have humanlike performance on telling apart pictures of cats and dogs that's something where sort of the test is can it emulate what we humans do not can it follow what the physical world does but again more complicated on the quest of the physical world is often what we want is not just the raw physical world but the physical world as perceived by humans in some way or another so I think we can kind of ask about all these different things about what's possible and what's not I think one of the things that was obviously the big surprise of a bit more than a year ago now was the whole chat gbt story of gosh you can have an AI that produces kind of pretty humanlike language and sort of why does that work how does it work and I think the realization is it probably means that humanlike language isn't as kind of complicated as we thought it was it has more regularities that we perhaps should be embarrassed for not having noticed in the past and and I think that same thing can be said about a bunch of the places where sort of there's been success in where where the the AI has been able to sort of jump ahead and do things that seemed magical to us it's like well that's because we didn't notice something that we could have noticed that that human brains could have dealt with we just didn't happen to go in that direction so you know there are there are all sorts of questions about um uh well all sorts of questions that that raises like one one question is okay so if the if the AI notices these sort of regularities in the world that we didn't notice in a sense science is about recognizing those regularities so is it possible that the AI can sort of uh find regularities that will kind of build drive forward science here's a couple of issues with that one issue is it's one thing to say yes I know what's going to happen I have this black Black Box model that I sort of Feed the input into a neural net and it says it's going to the answer is going to be three or something that's sort of case number one where it's kind of a blackbox thing case number two is I can tell kind of a human narrative or I can derive it's going to be the number three with all kinds of human steps inside there and I think it's uh there are plenty of cases particularly in technology where you know getting that camera to focus correctly on the on on the object of interest in the image it doesn't matter what the whether there's an explainable algorithm for doing that that's not it's something where once trying to achieve a technological purpose that's that's what we want it's I think less satisfactory in some cases if you say well you know why does this if if if you're trying to have science be this kind of um this this way of kind of taking what I what science at some level natural science at some fundamental level is is taking kind of the complexity of the world and figuring out how to make a narrative based on that that sort of addresses the main things that are going on in a way that sort of fits in our finite Minds that's kind of the the key activity of natural science is to do that and in a sense by saying well we just have this black box which can say this is what's going to happen but there's never going to be a human level narrative about about how that works isn't our traditional view of how natural science is supposed to be set up doesn't mean it's not useful doesn't mean it's not but but it's different from from what's been done in kind of Natural Science as it's evolved now it has to be said that Natural Science as its evolved kind of in a sense is is evading computational irreducibility wherever you say well you just have to run the system and see what it does that's precisely saying there isn't a narrative that that can be sort of a reduced narrative that we get to stuff in our brains that tells us what's going to happen so I think this this question about sort of when when can we expect that sort of you can just learn from experience and see what's going to happen and whether you can do that for systems in general or for the aspects of systems that we humans care about those are lots of different kinds of questions and another another isue is is is this maybe it's the case that if you're using something like a neural net and you're trying to take all of this uh sort of data about what happens in the world and you're effectively making a model for that you know in a in a in a vastly simpler case when you make a model for something you might fit a mathematical formula to some data and you might say that's our model in the case of a neural net you might be fitting in the end you are fitting a mathematical formula to the data but it's a formula with a 100 billion parameters in it or something and it's a formula that has a particular structure that corresponds to this neuronet and and there kind of the question of well if you have a model of that kind sort of how big a model do you need to be able to fit this or that kind of thing so for example let's say we we're talking about mathematical formula models that's often the domain of Statistics it's kind of like oh is it is this thing fit by two gussian distributions or something you know a formula that involves let's say you know four parameters or something like that that is is that a decent fit to what we're seeing you know can we say I don't know I don't know why I'm thinking of this case but the law of friction for example you know f equals uh the MU * n you know it's a linear law the frictional force is linearly dependent on the normal force that's a very simple model it's not precisely correct but it's a good approximation in many for many purposes and the fact that that whole collection of sort of physical phenomena is modelable with that simple linear formula is a is is an interesting fact now the question is if you say well let's take human language for example how big a formula do we need effectively to do predictions to model human language as human language is is uh um is is typically used and the answer we know from chat gbt is with 200 billion parameters we can do a reasonable job of of sort of essay writing type human language activities and there's sort of a question of of is that how does that that 200 billion parameters or something how does that uh kind of relate to what human language is and for instance could it be the case that there is sort of a a a point of Threshold at which you suddenly start to be able to do human language like things oh if we we only had 10 billion parameters we'll never get there and and so on now there's a little bit of kind of um circumstantial evidence about these kinds of things so if you look at the history of machine learning for example there have been these particular times when sort of major advances were made that sort of vaguely corresponded to the amount of training time the size of the neuron Nets these kinds of things the the sort of the early efforts in doing you know can we recognize uh you know digits from 0 to 9 uh that was you know 1980s and so on even earlier than that can we recognize you know 5,000 different kinds of images that was 2011 2012 um what you know they're these thresholds that seem to be somewhat associated with the sizes of networks and the amounts of training data and so on and so one of the things that's obviously an interesting speculation is if you want to get to the point of having reasonable human language like stuff there's a threshold I I tend to think the threshold is considerably lower than the 200 billion parameters because that happens to sort of package in a lot of detailed knowledge about the world that you don't really need to have good sort of Common Sense uh kind of linguistic capability but maybe it's a tenth of size maybe it's uh you know 50th the size it's it's probably not uh a million parameters that's probably not not enough so then the question would be if we look at kind of the history of biology and we say okay you know the great achievement of our species from a couple hundred thousand years ago or whatever was we got human language and notably another great feature of our species we got a bunch more brain a bunch more neurons in our brains and more structure in our brains and you know is it is it is it conceivable that there is a threshold at which you can have sort of reasonable human language I don't know I think that you could ask the question you know typical human languages have you know 30,000 40,000 uh words in them that that get used I think that's true of of all human languages I'm not sure um I think that's true of the the what 7,000 or so extent human languages I think that's true of of all of them uh now there are sort of constructed languages that have many fewer words but there are fewer things you can say about the world with those with those constructed languages you can even have ones down to a couple hundred words and you can have a kind of General Pleasant discussion with those languages but you can't have a I mean for example the the conversation we're having right now would not really be accessible to a language with 200 words in it now you know it raises the question of if we scaled up if we had a language with a million words could we express a lot more kinds of things than we can express in our language with 40,000 words or whatever could we you know could it be the case that in so far as we might talk about philosophy with our 40,000 word language if we had a million word language that we could have a completely different level of of discourse about things well I don't think we know but you know there is this this possibility that for example getting to kind of reasonable human language requires some threshold and maybe there's another threshold higher up where we get to something sort of a sort of superhuman language you could argue both ways on this I think one argument against it would be the phenomenon the principle of computational equivalence phenomenon of universal computation and so on that's a place where you want to if you want to sort of get something that can do arbitrary computation you say well I've got to add all these capabilities to the system and I've got to make a you know a touring machine you know we we know from my efforts and and so on from the mid 2000s and and um this this prize that we had for you know is this touring machine the simplest possible Universal touring machine answer yes we know that the simplest possible uh sort of touring machine that can do arbitrary computation is this absolutely tiny thing uh with two states and three colors so but sort of there's this question of once you've got to the point where you can do Universal computation then there's no further sort of fundamental computational sophistication you can get beyond that and in fact you know it could be the case that if you're trying to do these things let's say with an interet that's incredibly inefficient and that theet has to just go sort of you know do all sorts of handstands and so on to uh you know and and uh Cascades of handstands so to speak to be able to get to the point of doing what if you represented in a slightly different way would be a rather straightforward piece of structure so to speak but that you know but that so the argument would be once the neuron net by all kinds of stuff has been able to get to this sort of universal computation point then computationally it's it's in principle able to go further actually neural Nets aren't good Universal computers in the way they're usually used because there's this kind of one of the features of universal computers is they can kind of run with arbitr much m and for arbitrarily long and that the typical architecture of NE that's right now at least doesn't doesn't have that capability but so the question would be the phenomenon we see there is that with universal computation we get to this threshold and once you're above this threshold there's no further you can go and it could be the case that with things like language human language and so on that the same thing is true once you're above that threshold you can yes you can add more words to the language and you can get more specialized you Cam more jargon and so on but fundamentally you're there in terms of being able to have sort of an expression of thoughts that has the sort of this compositionality of thoughts can Nest within other thoughts and so on and that you're kind of uh that there's no further you can go or it could be that you know if we had a million words we'd be able to sort of get to very different places I think what we can see is uh there's there's a question of quantitative versus qualitative change what we can see is that as we kind of formalize things with for example computational language World from language effort so to speak as we try and sort of formalize the world in computational terms which has been our like last four decades effort to try and represent the world in this kind of computational way that it is certainly something as a practical matter where you can build much taller towers with that kind of formalized structure than you can if you're kind of just um uh sort of um uh just just using sort of plain language and so on it allows you to to know that your Tower is not going to fall over you don't you know the the foundations are solid the bricks are are hard so to speak as as you as you build the thing up so you know I think uh another thing to to sort of Wonder well there's all these questions about sort of when we look at the future of let's say science you know to what extent is AI going to and this kind of neuronet idea and so on to what extent is that going to change uh provide change to to what's possible in science it's worth saying there are things that are the the the whole using computation to represent and formalize ideas that turn into science that's clearly a winning thing and that has you know that has a long way to run so to speak and there are pieces to that once you formalize things computationally there are kinds of things you can do that you couldn't do before like you could say something like let's enumerate all possible theories of this let's enumerate all possible systems that do this or that that that have this or that structure this is the area that I like to call rology the study of simple rules let's say and what they could possibly do or equivalently let's enumerate all possible chemical structures and see whether any of them do this or that thing let's enumerate all possible algorithms of this type let's see whether any of them do the thing we want that idea of kind of rology and exploring the computational universe it's not a very aiish thing because it's mostly it's a thing where you're really just Computing out there but um perhaps then the question is can AI help you to figure out when you're searching this big space of possibilities can it help you to say that's a more fertile area than that and so on that's one thing it could do another thing it could do is be able to say once you've got a candidate thing oh I can predict more quickly that that's hopeless or that's interesting to look at I mean this type of thing you know there are endless efforts to do drug design for example using sort of aiish methods where you're enumerating and sometimes you're enumerating sometimes you're you're selecting I don't think I mean you could you can use the um the AI in a sense it's very easy to get the AI to be creative because you know you just throw a random number and you say let's put the molecules in this put the atoms in this configuration and let's see whether that's a candidate now the question is are you are you are you arranging it so that whichever way the dice comes down you'll always get a viable C candate or at least a plausibly viable candidate and for that it matters sort of what the underlying model of what these molecules might be like is and if you set the parameters up correctly then whatever however you pick them you're going to get a reasonable molecule and that's a sort of aiish thing to do to kind of figure out what's a reasonable parameterization so that you can then drive it with rology effectively um and and just start enumerating things so you know I think there there are questions like that I mean I think another type of question is if you're solving math problems like there was recent announcement about geometry problems doing geometry proofs um if you uh sort of if you say well let's you know is there a point at which human geometry proofs are kind of yeah we kind of know how to do those because we've seen enough examples that we can kind of piece the pieces together and for the level of things that are human relevant we can just say yes you know we we've seen enough examples we can just see what to do I think the uh the question of sort of can you always do well geometry is a special case because geometry is a decidable mathematical Theory ukan geometry is which means in principle you can just for example turn it into algebra problem we can do that in W language and Mathematica you can sort of turn it into algebra problem you just grind it and you can get to a result in a in a it might take a while but you know you can do it there are plenty of other places in mathematics even in arithmetic you say I've got this equation that involves just integers and it is you know X Cub + y cubal z Cub plus whatever and say find X Y and Z so that satisfy this equation there is no upper bound uh in General on on how far you have to look to do that it's a that's a undecidable theory and and so that's a thing where you can't guarantee if if you say well uh you know is is it conceivable that I could just sort of close this off by finding enough sort of rules for how to do it and that then I'd always be able to do it in limited time the answer is no you you can't do that so but I mean there's this question of if you want to achieve sort of human level performance on some particular task to to what extent sort of how how many parameters of uh do you need to be able to do that how much training material do you need to be able to do that and so on and to what extent are there kind of thresholds of uh of oh it doesn't work at all versus it's sort of achieved human level or something um and you know you might have thought this was a surprise with things like chat gbt you might have thought that as you kind of added more neurons or whatever else added more training data the thing would progressively get better but in fact that's not our experience in fact our experience is it was really crummy until it reached above a threshold and then it looked really good now is that a phenomenon of the thing itself or is that more a phenomenon of our perception I mean it's kind of like you could say I don't know this is uh uh what's a good example I mean you know when something kind of is good enough enough that we can really use it versus it's um uh it's kind of a little bit too crummy uh what is a good example of that there must be plenty in in sort of mechanical kinds of things I don't know if you have a um uh uh an automatic screwdriver or something where the thing doesn't quite have you know enough uh you know can't quite exert enough torque to be able to you know screw that screw SC in or something it's like well that was nice but it doesn't quite make it for the purposes I have because typical screws in a piece of wood or something require this amount of force and it doesn't really quite make it and so I might as well just throw that device away um and as soon as it's above the critical torque level then it's like okay great it's going to work and and it could be the case that that you know we we don't really have a good sense for something like language of what that kind of what that thresholding looks like my own guess is that there are certain once you've achieved once you've leared certain regularities you're above that threshold like for example I could be below the threshold where I knew basic grammar and I could be continually making up sentences which are just grammatically wrong you know cat dog chase Mouse I don't know cat dog red big was okay it's it's uh you know we can in some language that that word order might be fine but but you know I can clearly make up sentences which you immediately know are wrong because they just violate standard syntactic grammar in English for example and there will be a point at which that no longer happens where it's syntactically correct um and uh uh that you know whether and how that threshold works is is not clear you know I think I was yaking on here about a lot of things and not really even addressing the question that was being asked which was I think about um feeding sort of the next level of AI uh things that were the experience of the previous level of AI and you know I think that is both an opportunity and a danger so to speak I mean in the sense that right now we are getting AIS that do human relevant things because we train them on human produced stuff by the time we're training them on AI produced stuff we've got this sort of strange uh sort of Dynamics where you know the AI thing is if the a thing was exactly a perfect model for humans then training on AI based stuff will be equivalent to training on humans and it's all good but assume that the AI based stuff isn't quite equivalent to humans for example let's say we've got essays and for whatever reason intrinsic or kind of control the essays of very Bland that the AI produces and then the next generation of AIS are trained only on very Bland essays well obviously then over the course of time as we train more and more levels of AIS things just get blander and blander and yeah they won't say anything wrong they might say almost nothing they might just have platitudes so to speak but um but that would be sort of the the direction so I think it's probably true that anytime One is using some sort of one is modeling what the original sort of underlying training data was one is making one sort of acting as some kind of modeler Observer whatever of that original training data the the the kinds of things that you're doing to to be able to model those things are sort of reducing that training data it's not the whole training data you've made decisions you've you've got what amount to biases uh not biases of of social or ethical relevance necessary neily they might be but just ones that the model was of this kind and so it tended to do this or that thing and then you amplify that by going to more and more levels and then it's more about what was yeah I think this is the point it's more about what was in the model than it was what was in the original data in other words if you iterate many levels of you know you're going to train from the from the results of the model that was trained from the results of the model and so on pretty soon your reflecting the model more so than reflecting the underlying training data or possibly reality now you know you can imagine that in sort of the the uh the training of AIS let's say you're trying to get them to train on what the physical world is like maybe they're robots maybe they can go and make measurements about the physical world maybe they've got grounding in reality in the real physical world that they can go and and and mine more training data effect i l from the physical world it's us humans have only produced a certain amount of stuff you know you could imagine some some terrible scenario where sort of the AIS have decided that they just need more training data from the humans and in some kind of strange inverse Matrix type setup you know the AIS have just hooked up a bunch of brains and the AI just sort of poking those brains just saying we want more training data we want to see what do this brain do in that scenario what does it do in that scenario let's go explore all these different things and you have this sort of brain bank that is that is driving that is providing the ground truth so to speak for the AIS as the AIS might have ground truth provided by going and exploring the physical world for example so that's a that's a a rather awful scenario I think from our point of view now I mean you might say gosh you know if AI is going to be the next evolutionary level in the development of uh our sort of well I don't know whether it's a biosphere but but um the development of uh uh of of sort of the future of us or something that gosh you just want the AI to be as well trained as possible to carry as much of kind of the the human achievement as can be carried into sort of the the future of of what exists um I think it's uh um I don't know as as a as just a simple human myself I I tend to be rather uh uh human- centered so to speak and and rather uh not uh you know the idea of of and I suppose it relates to a lot of sort of in a sense both per personal and societal purpose so to speak is it you know does one think it's okay you know if the if a civilization wrote down all its great discoveries and they were passed down through the ages but the civilization was exterminated uh you know is that is that okay or does one feel bad about that I I kind of think as sort of a as a human one feels bad about that um as a sort of looked at from the outside whatever that means because I don't think it means much you could say look you've got more knowledge more abstraction more abstract kinds of things that were produced even though sorry the actual humans that produced it are long gone um you know it really depends on and it sort of it quickly devolves into questions and this is going to go sort of deep philosophy or something into questions about sort of the purpose of the universe and if it has one which I don't think it meaningfully can be said to have one um but uh you know this question of whether whether it's sort of good enough to have these kind of engrams of of abstract achievement or whether what matters is the details of the actual humans and I think the uh uh sort of it quickly devolves into meaninglessness if you say that what is going to be left is just the engram so to speak without the actual humans and their experiences you know I think it's uh yeah so that um uh yeah so I mean this question about if you if you let the AIS train the AIS and so on my claim would be actually sort of interesting inter in interesting thing because okay in in biology there are sort of forces of natural selection that are informed by the world as it is you know if the climate gets warmer and things like that there will be changes to the sort of forcing function of what is a successful species and so the species through natural selection will tend to evolve sort of fixed to the ground truth of the actual physical environment of the earth uh for now at least until we get species on Mars evolving and things like this or on the atmosphere of Venus that's a more interesting case I would say um but in any case we can imagine some interesting kind of uh natural artificial selection of let's let's make microbes that live in the in the Clouds of Venus and so on and perhaps do interesting things or things we think are interesting to them but okay so in biological evolution there's sort of a ground truth from the uh the physical environment there's also a certain set of constraints associated with kind of the way biological organisms are put together you know unless you're going to jump away from proteins and things it's we're fixed we we have a certain set of constraints about what can happen unless we you know it's pretty hard for organisms to make Wheels I mean bacteria have little fella at the back and so on but fundamentally you don't see a lot of big organ isms running around on or Wheeling around on Wheels uh because that isn't something it's particularly easy to make with you know blood supplies and nerves and all this kind of thing so it's um uh there are these constraints about the construction of the organism and there are constraints that are imposed by sort of the environment now it could be the case that being a wheel organism would be a fantastic thing for some environment that that arises but you probably won't get there because it's just not something that can be constructed from the components you have but I think an interesting thing with sort of the AI story here is if you say train from the train you know train train train in a big in a big train of training so to speak that you um will end up with something which is much more dependent on the underlying constraints of the construction of the AI than on the external environment if you keep going back to the external environment and saying what can we deduc from sort of the ground truth of the external environment that's more like what has happened in biological evolution so far I mean by the way we can imagine a similar situation in biological evolution if we're if we're evolving you know in a Petri dish or maybe even in in tumors and humans I'm not sure where where sort of it's in an environment where the environment is sort of fixed and you don't get to sort of go back and ask more about the environment and it's that that then it's more something which is strained by the intrinsic nature of the thing that's evolving so to speak rather than by what's sort of imposed by the environment but so my guess would be if you go the sort of the chain of of trainings of of AIS that in the end it will be more and more evident that oh you using a Transformer with this and that and the other characteristics in the in the in the system now you can say well how about if you evolve the the the architecture of the neur net sure you can do that I think then you will be recursing again to saying well what was the evolver of the architecture of the neural net and so on I think unless you you get to sort of external ground truth I think that you are inevitably going to evolve it the interesting thing to actually study in detail to evolve to something which tends to be uh kind of uh uh which which tends to reflect the sort of inner construction mechanism okay that was a super long answer to that let's see uh well okay let's there's a whole bunch of questions and comments about AI here let me try and address some of these but X asks do you think anyone will solve the rean hypothesis in your lifetime let's let's talk about that for a minute and we'll talk about whether the someone might be an AI what that would mean etc etc etc reminder about the ran hyp hthis formulated around what 1860 or so by Bernard rean the uh rean hypothesis is about this thing called the rean zeta function and as it's stated it's are all of the zeros places where Zeta of of uh s is equal to zero are all of those at places where on the critical line um uh s equals a half plus I times a times an arbitary variable where I is squ minus one so that's that's what the the thing is it's stated in terms of this sort of mathematical construct what it really is about is the randomness of prime numbers it's really something that relates to the density of prime numbers and really the rean hypothesis asks is there a certain degree of Randomness in the distribution of primes if there is that then the re hypothesis is true if there isn't that then there's certain regularities that we didn't think existed that um uh with within the Primes and um I think the when you look at the zeta function you sort of plot it out and you see it wiggling it Wiggles a lot and the ran hypothesis is equivalent to is when it looks at these Wiggles is there ever a a minimum wiggle a lowest point that's above the axis that's above value equals zero okay so you when you see it wiggling you realize this might be a hard thing to figure out because there's a lot of complicated Wiggles there now the real question though is what would it mean to figure it out you know it really relates to questions about what is mathematics and the the kind of last century or so view has been mathematics is based on you set down axioms you say these are the things that I assume are true you know uclid said that these are the things I assume are true now let me as a purely sort of mechanical process see what I can derive from that now that's a tricky thing because what do you assume is true like uclid for example assumed his fifth postulate assumed that no parallel lines can ever cross and that postulate turns out to be well it's a fine thing to assume you can do geometry on that basis as a as a constructive kind of thing but it turns out it's not true for the physical Universe in the physical universe space can be curved that was sort of the discovery of general relativity and uh that um that means that uet's axioms are just sort of this bubble out there that's not related as such to physical reality and you can invent different axioms for mathematics and you can say well somebody could say well I've got the rean hypothesis and I'm going to choose these axioms for mathematics that include very obscure things about infinite sets and how they work and so on and it's like okay with these AXS of mathematics Kaboom we proved the rean hypothesis you say well you know that was or for example worst case in in areas of theoretical computer science people sometimes assume the rean hypothesis so you could say well I'm going to make an axiim system and I'm going to get something which is you know almost assuming the rean hypothesis then if I prove the rean hypothesis like who cares that was obvious so the real question about proving the rean hypothesis is you've got to go from some axioms to prove the hypothesis now it could be the case that the axium that you typically use in mathematics for let's say arithmetic or something that those AXS are just not powerful enough to prove or disprove the reman hypothesis it could be the case that the reman hypothesis is something where you know you the the it's a thing where you could with those axioms alone you just don't reach that particular theorem roughly it's a little bit more complicated with reman hypothesis because it has the feature that if you can show that it is undecidable then it is in some sense true uh because of just the way that it's stated in terms of for alls and their exists and things like this but but roughly it's it's fair to say that sort of it depends what uh you know you have to ask given this particular sort of low-level formalization of mathematics and by the way most mathematics that's done is not done down to its lowest level formalization for example you know when format's Last Theorem was proved you know that the chap who proved it I asked him he said what aims did you assume he says I have no idea you know I just did it using sort of common mathematics of the time which is what everybody thought was the interesting thing to do to say uh kind of what would be the um uh you know to say Can you can you just go mechanically from certain axioms to prove that theorem that's sort of a different level of question not one of such interest to most mathematicians I mean there are efforts in modern times to make proof assistance computational systems where you're kind of fitting together puzzle pieces and you're saying you know if you can fit this puzzle this way then you have definitively proved this thing but it's a it's a really hard road to um uh to be able to do that and then even when you've done it it's not clear what you've got because the it could be the case that so so what you have to claim is that something like the hypothesis format's Last Theorem that there's a gaggle of possible Axiom systems for which that that a re reasonable axim system makes the thing true and that you could that if you tweaked the axim system so that it wasn't true lots of other stuff would fall apart but that's a little bit of a trickier thing now it could be the case that with the kind of standard level mathematics that don't go all the way down to these axim systems that you could prove the ran hypothesis and I don't have a good sense my my own guess would be that you need to go down to this infrastructure level that it that it if in the end it has a dependence on some infrastructure kinds of things I will say about the ran hypothesis there is a nice correspondence um there is there are sort of fairly simple to State mathematical problems like the 3n plus one problem you take a number if it's even you divide it by two if it's odd you say 3 n + 1 and then you ask the question if you start off with any number does that thing eventually reach sort of small numbers it goes to one and four and things like this or does it sort of eventually go off and get bigger and bigger if it always hits the 3 n plus1 it's going to get bigger and bigger if it hits the N over two it'll go down to something smaller and it bounces around you can make these plots they bounce around all over the place so it turns out there's a formulation of the rean hypothesis as partly due to well most of the real work um done by a chap called Yuri Mattis savich long-term user of technology uh who's been involved in proving lots of lots of interesting things about uh kind of what's decidable and what's not um but anyway he he kind of made a a version of the r hypothesis which can be stated in terms of these sort of iterations involving inges I tried to clean it up a bit got a simpler version um which actually see I think seemed to people some point so bizarre that um uh you know how could this possibly be be right but um but I think it is um and it's just you know it's a one line of w language code and just says if this iteration n goes to blah blah blah um if that iteration doesn't uh if that iteration always terminates then the reman hypothesis is true and so you've you've reduced it to this problem about sort of iteration of integers and that kind of makes it a little bit more kind of it brings it a little bit closer to these kinds of questions that are more about kind of going down to the low levels of axioms and a little little bit less about kind of can we construct this nice geometrical picture that allows us to use sort of traditional mathematics to go and um uh to go and work through uh different kinds of things so I think the answer is my guess would be that the r hypothesis might be proved to be uh undecidable from certain ACC Sims that's a thing that might happen uh you know fairly soon to know kind of the narrative story uh you know from sort of some common idea of mathematics that isn't as such the axioms I think that's a more challenging thing let's see uh oh was there's a there's a question here from Jamie will AIS um okay let me take a quick quick comment here from RBS aren't there thousands and thousands of Papers written that assume the reman hypothesis is true yes absolutely there are and um and uh let's see I mean the the status of those many of them are algorithmic papers that say this algorithm will be correct this randomized algorithm will be correct if whatever and and my guess is I haven't really thought this through properly but but my guess is that this whole question about what aims you're assuming to get the result H that probably relates to particularly when it's randomized algorithms that probably Tangles itself up with where did the randomness come from that you say well I pick these things at random and with overwhelming probability probability that approach is one this or that will happen that somehow the atiz of that random randomness will be sort of entangled with the aimation you're using to prove the hypothesis that that's that would my guess about how that will work um let's see Jamie's asking will AIS be the ones to explore space well I mean it depends I think that um uh you know in space is very not very human friendly it's the time scales you know the speed of light compared to and the distances between Stars compared to human lifespans it doesn't look good um and uh uh it's you know you want to get to the nearby star it's going to take you many current human lifetimes to do that and so well what do you do well you can sort of dehumanize the humans at some level you can say okay let's let's you know make the thing that's worked in science fiction forever but we don't yet know how to make work in practice you know cryonic suspension so to speak let's let's make that work and then sort of we we kind of uh uh you know we we freeze the humans and then they wake up in 50,000 years and they're at alpha centur or something um and uh you know in a sense that's a uh uh you know in that sense yes the humans can explore space if if we manage to make that technology work which my guess is one will be able to I think it will be non-trivial to have the you know the Interstellar dust part interplanetary yeah Interstellar dust particles that you know hit the spacecraft not you know over the course of 50,000 years or whatever not destroy any material we put in place but let's assume that's also solvable let's assume there's little robots that go fix things and things but in a sense it's like well yeah we got a human TFA centur in 50,000 years and and so what so to speak and we could also say well we got some kind of AI to Alpha centor in 50,000 years and again it might be a so what because it's it's um uh the you know this relates to sort of again to these things we were talking about before about you know you can create a thing computationally you can do a thing but does it relate to anything that we humans actually care about I mean I think that uh it's worth remembering that the the the the one way that we get to kind of explore the universe without going to all that trouble is through things that go at the speed of light whether it's uh light itself whether it's gravitational waves whether it's maybe neutrinos that go at least really close to the speed of light um these are things where they uh kind of you know they there are ways to sort of see different parts of the universe without physically going there in other words how important is it to see with your own eyes Alpha Cantor versus how important is it versus can you if you get photons from there that we can receive here you know is that is that a thing one can do and then uh you know I think it's it's worth remembering it's kind of ironic that for a photon because it's going at the speed of light and time is infinitely dilated at the speed of light to a photon no time passes you know from the time the photon is emitted to the time the photon is absorbed that's you know just a moment there's no no time has passed so photons that were emitted right at the beginning of the universe or at the point where the universe became transparent 100,000 years or something after it began um where there are photons that are just sort of going through through coming towards us and event some Photon from the cosmic microwave background that lasted 14 billion years of our time and finally got absorbed by some detector you know it seems feels very sad that it lasted 14 billion years and then Splat we detected it in some in some uh uh detector of microwave radiation that to the photon no time will have passed to the photon if the photon was kind of uh if if we had some kind of assembly of things that was uh sort of if we could imagine a mind made of photons so to speak then no time would pass for that mind between the creation of the mind and the destruction of the mind so to speak at a time which for our minds is is is vastly later so you know I think there's this question of sort of we're going to explore the universe we're going to send out sort of AI probes they will kind of go out and about in the universe and they'll arrive places 100,000 years from now or something you know I don't know that's a that's a complicated issue what what how one feels about that I mean it's kind of like building the pyramids and putting things in there that were sort of intended for the afterlife but end up being relevant for 4,000 years later and you know how do you feel about that you know that process of putting things out there into into the Universe I have a number of friends who've been involved in sort of putting things out into the universe so we at least have planted our flag on the UN Universe even long after we might be around or or not around and so on so I think it's um uh but if if one's asking the question you know for example within our solar system uh you know will it be the case that it's um we can have robotic spacecraft going and exploring things surely the answer is yes um there is a certain sort of coolness factor to having the actual humans show up on mars or whatever else rather than just oh it's a Rob IC probe and it's sending data back and we can get you know even if we can get very high bandwidth data back from it it still feels cooler and feels more us related if there's you know a friendly human so to speak who went there rather than just a robotic probe but I think you know it's clearly vastly easier to have uh I mean you know it depends a bit on one's view about sort of the value of humans versus the value of robots versus whatever else I mean if you know in some uh to my mind very undesirable world where you know you say well humans aren't that valuable we can send a bunch of them off to Mars maybe some of them will make it and you know they're cheap so to speak you know that would be a different world than than fortunately than the one we seem to live in where it's uh you know where the calculation of what you know what's it plausible to do is is different I mean there are plenty of places in the world to today where it's like well we could have a factory robot do that but why bother because we can have a human do it sort of uh for the same price type thing and I think you know you can you can imagine the same kinds of things with respect to robotic probes versus humans and in so far as you put sort the value of of human prosperity in life for the astronauts or whoever else very very high um you know it makes it much more attractive to to send the robot out where yeah it's you know it's kind kind of just money if it blows up so to speak let's see uh yeah well IBS is commenting you know if if only if only we could just send okay if on Alpha centor we had a sort of friendly Cooperative whatever where we just send a light signal out out there and then a 3D printer starts up on Alpha centur and makes the thing that we want to make and then it goes and you know does its study of whatever we want to study and it sends back this wonderful picture of this totally alien thing that we don't recognize as anything like life or whatever else some complicated chemical process that we say huh that's that's kind of interesting you know I wonder what that means so to speak um yes but I I think um uh that that's um that's kind of a we don't have a way to do that you know it's an interesting question whether we could in a sense whenever we well okay right now we are taking passive information from for example exoplanets you know there's light from a star it maybe goes through the atmosphere of the exoplanet maybe we get some wonderful spectroscopic conclusion from that but we're not and we're far from being able to say let's shine a laser in the direction of that exoplanet wait eight years or something and see what comes back we're not in a position to do that the energetics of that don't work out even close right now I mean the the fundamental problem there is if we could keep a beam of light sort of perfectly columnated then we might be able to do that but we can't because of the phenomenon of defraction it's inevitable that it will spread out and we can't and so that idea that then we have to put a incredible power into it so that even with the sort of inverse Square LW of it spreading out we'll still get something reasonable back I mean to give an example of actually probably the the the sort of the furthest out active things like that are you send a a signal to a you know the the um uh which one is it Voyager is still in contact or Pioneer Voyager I think um Voyager yeah so you know you can send a signal out to the Voyager spacecraft and it will do something and you get data back that's kind of like we put the 3D printer on Alpha centur except it isn't even close to to Alpha centur yet but so the question would be is there a way without having without humans having put something at the place where you're having a receiver could you imagine a situation where there's a thing out in the distance so so to speak that could receive a light signal and do the the useful thing with it and then send the data back we don't know how to do that right now and but that's I mean one could imagine that and that would kind of that would definitely make it uh even less sort of attractive to send out that human-built piece of material stuff I mean in a sense what one's proposing doing is sending out human-built photons but rather than human-built physical objects with little pieces of metal and so on and them let's see uh you know I mean I have to say in our models of physics the to us right now photons look like featureless you've either got one or you don't in our models of physics photons are probably made of this kind of complicated structure made of this network that represents the structure of space and the structure of everything in the universe it's kind of like saying I've got an Eddie in my water and I can see that Eddie and it's just an Eddie but that Eddie is ultimately made of a bunch of molecules underneath and so one could imagine that sort of a photon right right now we only think the only thing that matters about a photon is its energy and its momentum that's the you know the only thing that distinguishes one Photon for another is its energy and its momentum there's no labeling of a photon beyond that and um well actually that's more in well okay that that's that's basically a fair statement the um it's a little complicated which is the photon and that that um you'll end up with two of them and they have different effects and so on and the question is they exist physically or do they exist in bronchial space and that's the whole story of quantum mechanics and so on so that gets complicated but in a first approximation sort of a photon is a photon and it's kind of like a black hole in current theories of physics it's kind of like a a black hole all you all that matters is essentially the same thing energy momentum and there's a few other parameters that that matter as well the spin and the charge and so on but but to the outside of the black hole that's all that matters inside the black hole it could be the case there was some whole civilization that was you know whose planet was crushed by some collapsing star and so on and that's inside the black hole but from the out side you don't see any of that and might be the same way with photons uh maybe that's there's no way of of sort of cracking open the nut of the photon so to speak and seeing what's inside in the case of black holes we think that quantum mechanics in some sense can crack open the nut and see what's inside um and and that that um and so this question of okay you've packaged things in a photon that's going at the speed of light and so on you know is it possible you could sort of crack open that nut and um uh and see more detail about it that's I mean it's faintly conceivable given our models of physics it's certainly something far away from what we know how to do at the present time um or even sort of uh it's it's um I mean it's this question of are all photons really the same um and you know right now we don't have any way to distinguish them um and uh the um but that's that will be one possibility that will be one bizarre possibility for a sort of additional channel of communication I I'm I I don't hold out that much hope for that but I think that's a conceivable thing I mean something people are always commenting on is when it comes to you know uh have we met the aliens so to speak and I I always think that's an a very poor question because the details of us humans if yes if there was something in the Galaxy that was just another you know humanoid like carbon life form type thing we meet them and we say yes you're a human Carbono humanoid carbon life form or something we recognize what it is but you know if what it is is something that involves chemical reactions involving manganes and uh niobium doing this and that and the other and producing very complicated patterns which to the inner experience of the M magnesi manganes niobium thingy is like look we have this amazing civilization but we look at it and we say that's some weird chemical reaction which we don't understand and who cares type thing so I tend to think that the the sort of the the problem of sort of alien finding is a problem that that there are many complicated things going on in the universe they're just in the terms of our physics project for example they're not close enough in Ral space to us that we can readily interpret them but in any case the the thing that's often sort of a a staple for science fiction and so on is imagine there's a new way to communicate in the universe which is basically saying imagine there's a new feature of the universe that we get to bring into our perception like you know we've had eyes since the trilobites or something I don't know whatever it is you know a solid few hundred million years uh you know life on Earth has had eyes um but life on Earth doesn't have neutrino detectors life on Earth doesn't have gravitational wave detectors we've built bu those things as pieces of Technology our species has built those things but that's something where we get to see a new feature of the universe when we do that and one sort of one might imagine okay well guess what if if only we had managed to tap into the you know the gravitational wave sort of internet then sort of then we would discover the aliens now in some sense things like that are true because the this this point that there's a lot of complicated stuff going on in the universe and what you notice depends on sort of how you can do the noticing and so that's something where you can expect yes you know once you know about gravitational waves and you can detect them and so on you'll be able to see complex processes going on in the universe that you weren't otherwise able to see and you could say well we've detected sort of the alien intelligence I I might say that you know the weather on Earth is an alien intelligence it doesn't happening very close to our intelligence but it is something that has the same kind of computational complexity that um our our brains have let's see it's a question from imagination here do you think Ray csw's longevity predictions are likely to happen in our own lifetime haven't talked to Ray in years actually um so I don't know I don't know how he's doing on the longevity scale um but uh I remember having lunch with him a number of years ago and um the uh was impressed at the at the number of pills of different kinds that um he he took along with one piece of lettuce or something so you know who knows maybe that's a winning solution I think that the um uh the question of what will happen with sort of human longevity is it's interesting one I mean you know we've been in many countries sort of the life expectancy has been increasing by about a third of a year per year for the last 70 years or so I think most of that is cardiovascular kinds of things that things which are essentially mechanical people can detect and so on I I tend to think and it's certainly my own personal sort of uh approach to this is the first step is detect things you know do all those tests make all those measurements you know if they all say it's all fine then you know assume it's going to be fine but by doing all those tests you get sort of an early warning of what's going on I mean it's kind of like how did airplanes or general anesthesia two examples where sort of there have been great advances in sort of their effectiveness and uh uh and not crashing and not not failing type type thing over the course of years and and I think the thing that's really been the the Big Driver of those is better senses better kind of awareness of what's going on you know is the taale of your airplane about to fall off well you know you might have some sensor that would tell you that rather than just you have to wait and see what happens so I think the um you know that that well like you know metal fatigue detection or something from from uh uh whatever but but different different kinds of kinds of things but the point is so I think one you know that that's surely a driver of uh of longevity but it's not going to solve the problem because that's just saying you know detect things early maybe you can patch them you can fix them if you detected earlier um that still is the case apparently that life eventually wears out now at some level you know there are plenty of organisms where kind of that organism buds off another piece and the bud kind kind of keeps going and the organisms gone on for a gazillion years the organism in some sense you could say but it's not the same organism but it isn't the same of us either because you know our cells are continually turning over okay there are some cells like nerve cells that don't turn over at least not at all quickly um but you know the the the is it us well you know our skin cells our gut cells whatever else blood cells they're all turning over fairly fairly rapidly and being replaced by new new cells and so you know I think there's this question of but but we think you know the part of us that is critical to the usess of us is mostly brain type things whether that's really the ultimate story is not so clear I mean I don't know how how much of the usess of us depends on things other than brains might depend on the endocrine system might depend on the immune system might depend on the gut microbiome who's to say you know those things you say well well those aren't really relevant but they might change for example your approach to life might be different depending on sort of uh you know hormone levels that are produced by something that has little to do with the brain as such so it it's not completely clear what the package that you need to sort of maintain the usess actually is but I think the um uh when it comes to you know maintaining the whole package that um you know supports the brain and all these other systems that's the question is can can you do that and sort of uh you know how will one increase longevity well the first question is why do we age to begin with and that's not completely clear I mean it's you know there are different sort of things that go wrong and there are telome that fall off the ends of DNA but then they can put them back on again you know think the enzymes can put them back on again then there are sort of oxidative damage things essentially just burning up and um there are and there are sort of deposits of this or that that get produced all kinds of things there are sort of tangling of the DNA there are all sorts of different things that that um uh that go wrong obviously they can be fixed because when you go to the Next Generation those things are restarted and it's not obvious that you can't take you know the things that you have right now and kind of restart them from stem cells for example uh it's it's I mean there are there's increasingly increasing success in kind of taking sort of well first thing was taking a cell that exists you know some random skin cell or something and reverting it back to something that will be a cell that can differentiate into anything just like our original uh you know fertilized egg or something cell can differentiate into all of us so to speak and so that's um uh you know it's it's clear that it is possible to turn the clock back because it happens with every successive generation so now the question is can you sort of turn the clock back within an organism you know in a sense can you rebuild the plane even as the plane is flying so to speak with with us organisms and we don't know yet um is it conceivable yes I think it's conceivable I mean people the very fact that was discovered what what is it 15 years ago 20 years ago of how to no 15 years ago I think of how to go back from a a sort of a a specific cell to something that is a stem cell that has the potentiality to turn into to lots of different kinds of cells that was only discovered fairly recently and I know there are experiments going on of can you kind of cure aging by just sort of having cells uh be be fresh cells in some way and you know there there are things where you say let's you know when there's cells that didn't make it and that one kind of became senescent you know when that when that cell was um uh sort of died off let's get rid of those cells and then maybe they'll be more cells get produced that that don't have those same aging characteristics I think um that uh uh you know but but it's also the case it's very strange thing that there is so many clocks in bi biology I mean the thing that always is is a little shocking to me is you use a facial feature recognizer like the one we have on W language and it'll tell you you know it takes picture of you and it'll tell you how old you are and it's like really you know and it gets it right quite accurately I mean I I was thinking for me for a while that I might have been in the training set for the for the machine learning classifier that figures out ages but but I don't think that's what's going on I think uh what is going on what's going on is that there are gradual changes in facial structure just as there are gradual changes you know eventually they by the time you're like 40 years old or something the plates of the skull fuse and so on there are things that happen over the course of many decades and which sort of inexorably happen and which by the way the AI can notice even though we humans don't so readily notice these things and so that's sort of an inexorable clock through life now that particular thing of of moving around facial features and so on and you know pieces of cartilage getting bigger and whatever else is is probably not relevant to whether you live or not but it's certainly a a kind of a a sort of a reminder of the fact that there is an intrinsic biological clock and what is needed to divert that biological clock we don't know I mean and of course with things like tumors the what's happening there is that sort of the usual of clocking and limitations on growth and so on that exist in biological organisms they have been they have been removed and we're kind of back to a form of life before we had sort of individual organisms that grew to fixed size it's just like just keep growing and that turns out not to be very compatible with uh you know with organisms that were at our stage in The evolutionary process so to speak so you know but what it takes to get into something which sort of grows like us but doesn't go out of control like that it's almost one of these undecidability problems you set up a certain set of conditions in your molecular biology is it going to be something which sort of does the right thing or is it going to do something computationally irreducible and unexpected and do the wrong thing and for example start sort of growing forever or some other bad thing like that but so you know I think look in the end there will be human immortality I I I'll be I don't think there's any fundamental reason to think that that's impossible what exactly it will mean and how much is sort of turned over in that process if it what it means is you kind of get to have a biological organism that sort of has the same memories and can remember from 300 years ago and so on um my guess is that will happen and uh how much of that will have been sort of oh and you by the way have a digital sort of assist that's connected to you know forget having uh actual organ X let's just have the the uh uh the kind of technological version of organ X why why bother to have an actual liver or something just have this kind of artificial one that um uh sort of plugs into the is plug compatible with the biological one and maybe much of what you have ends up being that that um uh artificial technological solution and then you have to ask what does it mean to have human immortality in the case where almost everything has been replaced by something that isn't really terribly human and of course that gets even more complicated when you're replacing pieces of brains by sort of the the digital assist so to speak or the neural net the artificial neural net that sort of oh yes well you know you're something went wrong with some piece of the hypothalamus so we replaced it with some artificial hypothalamus and then you know and then it's like how far does that go and at what point do you have human immortality versus essentially uh just or I say just but it might be uh you know that this this um quotes artificial this technological immortality you know I'm certainly hoping as a as a person who's getting old myself I certainly hope that uh the advances that might happen that would get one to the point where you know the the life expectancy is increasing at more than more than one year per year so to speak so that you so that it sort of goes to Infinity um that that happens uh sort of soon enough to not uh you know to to not sort of exit oneself it would be great I mean you know the other the other opportunity is the whole cryonic story and the ability to sort of have a pause in uh and sort of freeze everything until for a better day so to speak I I think the challenge with all of these things the challenge with that one is the discontinuity of experience may you know what what as you know if one's been around and exposed to the way that technology is evolving and so on we're not that you know one isn't you know if if you took the me from 1965 or something and brought me to today there'd be lots of things where I just I have no idea how this works what the heck is this um it's uh you know the world has changed because of Technology primarily a little bit because of social kinds of Dynamics as well but um uh one might sort of say some things that will be terrible today or say some things that will be Charming today from you know jumping from 1965 or something but I think that uh you know that Dynamic of um of kind of you know if you are continually around and sort of paying attention to the the development of the world you're going to to be sort of still in your time if you kind of were frozen for 100 years and then come back then you're not in your time and it's not completely clear how that plays out let's see sumof comments would it make sense for alien AI civilizations to broadcast radio or other signals with information about how to build an AI that can propagate a universe faster than materal space travel yes but you know the thing that's so difficult is it's so hard to go outside of one's current human purpose you know the things that have been sort of you know bashed into us by our struggle for life in biological evolution etc etc etc it's so hard to kind of go to the the kind of the abstract purpose of you know now you know it has to be said that if your if whatever you do causes you to not exist then there's nothing more to say so there is a certain sense in which sort of the propagation of something and making more of it exist is a self-fulfilling thing that has to happen so you know I suppose that's an argument for why the I mean you know one could say that um let's see what will be an extreme thing to say one could say that um look you know all of the um what could I say something like all of the interstellar grains that have all these particular molecules on them that um we can um uh uh you know that that we get uh radio signals that show us that this molecule exists and a slightly weird molecule we never bothered to look at before exists gosh golly there must be a you know an Intergalactic conspiracy or at least Interstellar conspiracy that's putting these particular molecules on these grains that we will detect and that we'll say gosh that's an interesting molecule let's go make that molecule here now that isn't our usual interpretation our usual interpretation is it's just physics that's putting these molecules on these Interstellar grains but I think that this it's just physics argument is a complicated one because look at what our brains do it's just physics it's physics where we say it's packaged in a way that's very special to us etc etc Etc but at some level it's just physics all right maybe one more question or comment uh bunch of comments here about um uh diminishing returns Fu says related to AI machine learning um H comments AI will create this time of Step function upwards and everything um but that may not be the case due to diminishing returns yeah well I think it's an interesting question to what extent you know what we've seen sort of observationally in the Technologies we've seen you know the big jump in 2012 with with um identification and so on image processing slightly less heralded big jumps in speech to text for example big jumps and language generation and so on it seemed like they they keep on being these sort of jumps where something gets to the point where it wasn't possible and now it's possible and you know there's the question of well you there's a certain set of things we humans do and we can kind of check them off could we do that one could we do that one and so on and I think the uh uh and then there's the question well what about the things we humans don't do right now and that's where things get challenging because you know it's very easy to make a computational system that does a lot of stuff that we humans don't do problem is it's not clear we humans care about it I mean in my life I've spent quite a lot of it studying rology kind of the the study of simple rules and what they do kind of what's out there in the computational universe your average you know cell automaton with all these rules very simple rules often very complicated Behavior you look at it and I say that's pretty interesting I but I can't really humanize it very well I can't really say and that's why you know it makes my you know me have a better lunch type thing it's um it's much more abstracted from that and I think this question about once we've we we know there are certain things we humans do now there's the question of what else is out there in the computational universe and there's a lot out there in the computational universe but the the history of civilization the history of sort of intellectual history is about making these small moves away from what we've already done sort of building on the tower we've already built then we go to the next step and there's this sort of vast expanse out there of what's computationally possible and most of it we humans as of today would look at it and just s of shrug and say we don't care I mean I I think we care because I care about rology and studying this computational universe and sort of the the science of that but in general in term I would not be able to say that there sort of immediate everyday applications of what you see out in the computational universe except things that are pretty close to where we've already gone so far so I think this question of of sort of you know can the AIS do you know they go out into the computational universe what do they find I think the limiting factor there is mostly what do we humans care about that is we can and that becomes sort of the what are the humans going to do well the AIS can zoom off in any direction you know the AIS are you know fast driving cars or something but it's not particularly useful to have a fast driving car unless you have a you know unless there's a human driver who's deciding where the car should go I mean the disembodied fast driving car can zoom off in random Direction in the desert or something but it's it's like the humans are probably going to say who cares um the the you know the the thing the place where the humans fit in is in this you know so which direction do we want to drive in and I think that's something where when people wonder sort of how do humans fit in to the world of AIS yes the AIS can provide the fast driving cars but it's still got to be kind of one's got to have choices which are for us humans in so far as we are we care about how we what we perceive in the world world that has to be determined by what uh by kind of what we humans choose so to speak so I think um uh you know they're both practical technological things about these thresholds associated with different kinds of things you can do then there are these kind of so what does one care about doing that one gets into and uh you know I I do think it's still sort of an open question if in the what does one care about doing we care about for example making human language let's say there's another level of of thing there that is I don't know you know let's say you have sensory information not from just you and your eyes and your fingers and so on but you have sensory information in detail about the whole world now in a sense we already have that because you know we're looking at the web and so on and we're getting uh you know live streams from all over the place and all that kind of thing but let's say we were able to kind of have sensory information sensory input that really was looking at every you know Internet of Things sensor everywhere in the world and maybe there's a different level of kind of description that we can start to apply you know human language maybe we could say is relevant if what we're talking about is the kinds of things that we as sort of humans walking around and you know whatever you know about you know five and a half feet tall or whatever it is um you know that that um that we experience but there's a quite different level of thing that's kind of a super not language like but who knows what it is some some way of organizing things that's relevant if you are seeing every every sensor in the in in the world so to speak I mean it's kind of uh you might say there are certain kinds of things that are worth thinking about if you have eyes and you're getting so visual data different things if you only have touch for example all right I better go back to my day job here um but uh thank you for a lot of interesting questions and comments today and um got me to think about a bunch of things I haven't thought properly about before so thanks very much and uh look forward to chatting again another time bye for now
