[Music] most people don't know what current AI can and cannot do it's not just like this mythical Black Fox that they just have to be scared of we are all centors already we live most of our lives on digital technology connected with other human beings we are part of some weird form of a hype mind and we are all kind of cyborgs and it's like yay we invented
fire oh guess what one of the greatest things in the world it also is one of the most destructive things in the world more people are by vending machines and dpd4 driverless car crashed and killed somebody take them all off the roads they are absolutely not to be allowed well wait a minute compared to what do you know how many crashes killed
people with a human behind the wheel would you trust a government regulator to test software for bugs before you installed it on your computer it's like big brother is actually little brother who's not so terribly frighten we really litigating the same arguments in like a 10e cycle and it feels like we're stuck in Groundhog [Music] Day well hello everybody it's Jim o
Shany with another infinite Loops one of my favorite people and how many times have you been on Roy uh two three not enough good answer my guest today returning is Rohit Krishnan the proprietor of the strange Loop Canon substack but today we're going to talk about a great book that he wrote called building God he mistifying AI for decision makers so right like AI is
filled man they think Wall Street and other uh professions are are just built around buzzwords I mean like foundational concepts of AI are you know neural networks long shortterm uh memor scan models deep fakes maximizing objective functions lions and tigers and bears oh my so can you I'm gonna give you a lot of time to do this first question too can you try
to simplify and demystify AI for our highly intelligent but still needing to know listeners and viewers what's going on in AI and welcome thank you very much thank you for having me back um as always it's a pleasure to talk and as always you start up with a real softball so I appreciate that sure so um two thoughts that I started I
ended up writing the book accidentally because I ended up writing it to explain to someone after having a bunch of similar conversations about what they actually ought to know and two things struck me number one is that um most people don't know what current AI can and cannot do and that is a giant question and number two most people don't know how to think about the fact
that what it will will be able to do at least in the future so the nod the name of the book ended up being a nod to the usual discourse that exists fear and lothing and uh amazement at the fact that we are ending up building God so to me like the core question that I was kind of struggling with is when I was talking to folks across multiple
Industries whether it's in you know biotech or health or insurance or financial services how do you explain to people where they might actually be able to use um AI in order to kind of do their jobs better and what do they need to know about the history and where it came from so that it's not just like this mythical black box that they just have to be scared of or don't understand
and input all of their hopes and fears and project all of the worries onto and same holds for policy makers right I mean if you hear the discourse that actually exists a lot of it is fairly apocalyptic or at least extremely high level that starts at the point of you know they can reason they're getting better therefore they will get better at
everything therefore we're all going to die or some version thereof which exists in multiple lot spots so my contention is as sort of Highly simplifying is something as follows that we have been exploring how to encode information in various different formats through neural Nets and it's very before and after for a fairly long time it turns out um that if you do arrange
them in a particular form in a particular series and if you do matrix multiplication in a particular series along with a couple of other neat little tricks you can encode an enormously large amount of information in a substantially small Corpus and not just that it kind of preserves the internal structure what is the benefit of that it
means for the first time in human history we have computers that can actually talk back to us based on what we actually train train it on now whole list of questions pop out of that how well does it understand you know is it does it does the fact that it is able to talk back to us mean that it actually can Gro the meaning of the things that
we are asking how well it does it do out of distribution a lot of these questions are highly unclear but the fact of the matter is that at least as far as we know today what we train it on is actually what it is best at regurgitating back to us but this also means that based on this plus a huge amount of research that is ongoing at the very moment including a list of
things that I say like you know whether it's multimodality whether it's on better memory whether it's on you know differentiated structures not just transformers going to State space we can actually build far better model soon which will be able to understand a lot more about the world that we live in if you put it together if you're a decision
maker your decisions come down to something like is one of these AI is going to come and replace a large number of my workers either because you're worried for your job or because you want to actually increase your margin if you're a business owner and the answer to either of those questions is actually yes but not exactly in the way that
you're thinking because where the places that it starts is the places where the trading data is most easily available where reasoning is not hyper complex and you can actually replace part of the uh value chain relatively easily without replacing parts of the value chain on either end what that like you can replace the creation of an image but if
you want to actually put that image into a whole series of um workflow to create like a frame of a movie at the end this turns out to be a little bit more complicated we'll get there as well but that is the progression that I'm trying to imagine and in trying to place it in that kind of a um format the the hope is that you see that this is the continuation of a long series of
innovations that we have been making since we going back to the original discovery of algebra and therefore contextual Iz to say that like even though that sufficient degrees of change in scale creates degrees of change in scope there are specific areas where you might be able to apply it and actually create substantial change hopefully that
kind of sually covered it not quite suly but you did a much better job than than I was expecting uh that's a yeah man are you a large language model yourself Ro I feel like it sometimes man there are times when I speak and the tokens just come out you know all right so up next we have a question from and large language model itself I asked a large language model
what it would like to ask you and so I'm going to read it to you and hopefully have you answer it this large this large language model says and he's speaking to you or it is speaking to you in your description of AI as a fuzzy processor you acknowledge a level of unpredictability in AI Behavior how would you balance the need or
predictable AI systems with the inherent uncertainty of their fuzzy outputs in critical applications extremely important question uh thank you uh unnamed AI who uh for for for reading the book you know what's funny about this is like we spent a very long time in in in software specifically where we try to codify ways of working so that we don't have to redo
work on top of it um I have this uh thesis that software is one of the our software programs are one of the most complicated human achievements that we have ever made because if you think about the dependency graph of anything that we normally use like Zoom for example it is it is much longer than almost the dependency graph of anything
else that we have created in the world any physical for example it just you are all we all depending it's like the old XKCD cartoon of a small OSS project except that it's not one small osss project it's like 1,500 of them and small changes in that can destroy everything the reason that exists is because software once written is highly deterministic and barring the occasional
bit flips actually does what it's supposed to do however that creates a problem because that it only does what it's supposed to do if you know what to ask it right like if you ask it highly specifically it does what you wanted to but otherwise it can't and hence you know the need for programmers who actually know how to ask in questions in highly structured formats for the first
time now that we have created something that can actually talk back to us um in our own language or convert our uh vague extrapolations and hopes and dreams into something slightly more legible and structured we have something that I call a fuzzy processor which is that unlike the processes that we created which are highly deterministic mathematical
operations these uh deterministic mathematical operations somehow manage to give convert unstructured information to structured information and structured into unstructured but the drawback of a fuzzy processor and you are one and I am one is the fact that if you ask us to do the same thing twice the chances that we will do it exactly the same the way an
old fashion robot would have done it is not very high because you know this is why if you think about the most the professions with the most drudgery these are the ones where everyone is treated like a cog in a wheel and assembly line and marks and yada yada y because it is what is alienating us from our very own Humanity right you're kind of picking
and packing the same thing repeatedly you're like a machine you're not using your inherent um human Tendencies so with an with an llm the fact that it's a fuzzy processor means that you can now use it in a lot of different places where you could not have used any AI uh or any kind of software before because it can effectively be a replacement for parts
of sort of different jobs that people might actually be doing however the problem is that if you or I as fuzzy processors are used in those places we can be tested we can be evaluated like you know if I'm hiring someone for a job I know that they're not perfectly predictable however I can talk to them and kind of get a sense of how
unpredictable they are and like how they would actually deal with different situations and monitor those in different ways and ask for you know previous employers or references or interview them and find create basically this like um cone of uncertainty if you will like I can bound it so I can I know that they're not completely crazy I know
that they will do some things but it'll be within the bound of error so with llms and fuzzy processors we're at the early stages with that right I mean we have this um the inherent fuzziness is problematic only because you cannot depend on when and how it is actually likely to be fuzzy that it might end up going in any kind of random Direction so for us to be able to use it in any
actual real life situation especially in critical applications we would need to have a whole lot more confidence in how precisely it works we would need to like not it's in internals of like you know specific nodes and weights and stuff like that we already know it but it's slightly unhelpful it's like doing I don't know neuros to predict behaviorism
I don't think it is hugely valuable in and of itself however we do need to bound Its Behavior in some sense so that we know it cannot go completely um off the rails when you're trying to use it even with that I mean we we are speaking what after the latest boing disaster not not that long after so when you talk about complex systems where large number
of Parts actually interact with each other there are the the possibility of you know something going wrong always exists so the the way we solve it in real life is by having stringent QA and multiple checks and huge levels of evaluations and large amounts of redundancies and the exact same principle applies for fuzzy processors as well where the only way to make a
Fuzzy processor function in a critical system is by having large number of valuation so that it can bound it creating enough structure around it so that even if it does something weird or crazy you can actually you know cut off those particular probability branches of the tree and you can kind of direct towards something and having large
amount of redundancies so that you can actually ensure that the output that is coming from it is effectively usable so that even if it does something crazy or stupid the errors are not continuously compounded over a period of time um it's like that u i i i don't know whether this is apocryphal but I remember hearing the story about Elon where uh
they were trying to send um computers up along with the stalling satellites and obviously like radiation shielded computers are very heavy and highly expensive I and uh radiation shielding is important because bit flips are more common when there is higher levels of you know radiation that actually hits once they're above the atmosphere and I think his solution or the solution as
one of his engineers in that particular bu profl um story was to send three and they would just vot because chances of all three getting hit simultaneously are much slower and that's a way that we normally that's a way to use redundancy to solve for unpredictability and I feel like a similar kind of pieces has to exist with respect to sort of llms as
well and a COR of this is that it has to be local and it has to be fast because otherwise you can't do thousands of recurrent API calls very easily unless you can ensure that they will run ridiculously fast and you do need checks and balances um I know at least the last point is something you're highly sympathetic to so but happens to be the
truth indeed um and uh again and I asked this particular large language model to make your argument for you and you actually just hit on several of the things that your AI counterpart came up with he did get a bit it did get a bit cheeky though when he suggested that you Rie uh as a human might be seeing fuzzy from your perspective maybe you're the
fuzzy one said the large language model and I went oh snap yeah that is awesome I mean it's possible Right like parolia is like a known fault of uh you know as humans so it is it makes sense that we are the ones who are seeing it fuzzy and it makes also sense for it to think that just because like look ultimately if you kind of boil it down they are
doing deterministic number based Matrix multiplications and tokens which are also numbers which means that the output there cannot be highly probabilistic it is actually deterministic it seems like it is probabilistic to us or fuzzy to us because it doesn't conform to the same Notions of how we might actually go ahead and solve it which also makes
sense because there is no logical reason to believe that you know a large language model trained on an enormous Corpus of Internet you know um uh exhaust yeah yeah yeah there's no reason to believe that that will have the same world model sitting inside it as we do even if we were solely autor regressive large language models who behaved in a similar fashion so maybe they are just
like us just from you know it's like Marvel Universe they might be from a different Multiverse which is a decent way to think about it actually you know that's so funny I I went down that rabbit hole as well and it is a lot of fun and it generates a lot of great in my opinion sci-fi novels that actually don't have to be dystopic uh they
there's a lot of uh them that the turn out to being pretty pretty cool we programmed in this concept of er patterns which are Primitives basically you know what what large language model and other multimodal developers call you know the the AI itself in its on tuned on formatted form they call them Primitives and that that goes back to ER patterns which are kind of the basic
building blocks right and I had read something by Howard Bloom where he's saying which interesting about her patterns is that you know deterministic precise mathematical principles sort of the heart of deterministic thought can suddenly spring forth uh something that is much more complicated much more bounded by probabilistic
thought and and uh outcomes what do you think about that you know um I think I've written about a couple of these things before which is that at a sufficient degree of um complexity highly deterministic systems can also show highly indeterministic outcomes um I am by no means a first person right I mean this is a fairly it's a it's a common Trope in pretty
much anything to do with Chaos Theory or like even things like you know sand piles and Grains at a point of avalanche and Cascade and there's there's a bunch of these questions which are um in my opinion like more feasible to see happen than to predict how it will happen because prediction requires you to effectively run the experiment so to
speak um and I'm fascinated by that so I think like in some sense we we in like normal conversations quite often complicate um indeterministic with random and they are two diff or or unpredictable with random and they're two different kind of processes I mean like there is this um the common argument that people make against things like Free Will is like
you know everything is a physical phenomena physical phenomena given a sufficiently powerful computer might actually be able to get simulated and therefore you might be able to predict it and it's one of those things when you know logically it might hold true if and only if the computer that is predicting it does did not need to actually run the
simulation in order to predict it and if it did then from the perspective of the people being simulated Us in this instance the outcome will still end up looking indeterministic right unpredictable even though like the theoretically everything was uh as pre-ordained I know this is kind of vexed and driven more people mad than me but I think there is a core kernel of
Truth here that just because you can't create beautiful analog equations to predict the behavior of a particular piece of software physical phenomena whatever does not mean that That Is Random it just means that at a at a certain degree of complexity there are way more permutations and combinations of how things can go wrong than there's feasible for us to I don't
know conceivably identify and as we said in the previous like the only way to solve it is by having sufficient amount of QA in redundancy and kind of bound the system so that you can actually be relatively sure that it does what you want it to do I mean you know stock markets are a perfect example of this right I mean the the the flash crash is
my favorite example of this like you you know it's not an intended behavior of the system but it is one chaotic outcome that could have happened and how do you stop it you don't stop it by like stopping each individual trade or analyzing each one you stop it at the macro level saying like if it falls a little this much we got it off which is a which is a macro Behavior that then
controls the micro behavior of each individual algo which kind of takes that into account and even if it does hit we mean that the like the worst case scenario is bounded and you also covered that in your book because you posit that uh we could have a so-called flash crash uh of uh Ai and and why don't you tell our listeners a little bit about
your solution for that yeah I I mean I think the the reason I kind of CAU to that was because for for a few months I been thinking about what is the worst case scenario do you know what I mean like I mean the the the general discourse around the worst case scenarios from AI are to me seem like positing too much or at least kind of giving the it it the
frame of the question implies the answer as opposed to kind of the logic that underlies the question implying the answer and to me like the one the the one part where I do see a problem is that suppose we do create AI that is fantastic and it's relatively is steerable the way we are saying and we bound it and it actually works inside um any of the institutions we think as
important you know like whatever government and Military it's already kind of there and like Finance it's kind of it's going to start coming up more and more what is the actual outcome that is coming from it that is a unpredictable and B somewhat catastrophic and the the best example there is some form of like it is a flash crash effectively right because a large
number of algorithms working with each other against each other in tandem with each other as part of that ecosystem creates new syns that are um impossible to predict and therefore not super easy for us to kind of guard against and to me the solution is like it's the same one the only way to guard against it is at the macro level like you can't go
Solution by solution and say unless we can perfectly predict the outcome of this particular system we will let it go off and do what it wants to do cuz if you could perfectly predict the outcome of the system we didn't really need the system in the first place like it's it's kind of you know arguing against the premise of the question in the first
place so the only way to do that is by having macro conditions which are non-negotiable where we kind of create protections against in the financial services world this would be things like you know a macro condition that says like we don't like flash Crashers so we will prevent artificially the market from doing something crazy just because
of the com upop a large number of algorithm kind of talking to each other we will have to do something similar on the AI front as well where if you don't wanted to do certain outcomes in a particular system we have to go from outcome first rather than sort of Al go first you're not going to prevent that from by I don't know bounding the number
of flops because even with a lower number of flops we can find enough ways for it to kind of screw us up assuming there's enough number of them that actually interact with each other but the only way to stop that is step up a layer of aggregation actually stop it from the from creating the chaos that we don't actually want it to do yeah and that kind of leads us into uh the uh
various attempts to create legislation around AI I I was looking at the EU uh legislation and I thought of The Beatles song Mother Superior jump the gun uh because you know uh aren't we maybe overestimating current capabilities and underestimating complexity you know doesn't true AGI require multi-disciplinarian integration from a variety host of different
disciplines uh you know oh maybe like we're talking about AGI because that's the other thing that I'm noticing a lot of people I think this is a classic case of people who have almost no no understanding of what they're trying to regulate thinking that they understand it completely and then compounding the problem by at least it maybe it's just
me I'll get your opinion but it it seems to me like they have just already let's assume AGI is already here and uh let's regulate against that and you know then I just ask questions like you know wouldn't we maybe need a workable theory of conscious this in humans and other animals or seni and do you guys even and gals even know what like you are really
doing it seems to me should we be inflicting human OS which is great I love it we run it and it unfortunately doesn't have a users manual but you know we're walking around with quantum computers up here in our nogin and you know it even makes me think kind of like wienstein the limits of bil language are the limits of my world so so what are your thoughts on
that I I know in the book you also uh have a a suggestion for the appropriate way at regulation and and I think you'd probably agree with me that at least specifically looking at the eu's version that ain't it yeah I feel you know it's it's a it's a bit like a uh it's like a weird arms race where on the one hand people are saying that that is hugely important and hugely
powerful and therefore it ought to be regulated which is almost seen as a fade comply like it's it's it's like an assumed um outcome and then there's a bunch of other folks saying like of course it has to be because the the uh cost of inaction is so high that we have to kind of regulate it now what's interesting here is that like the word
regulation is used interchangeably in both of them but as we both know like regulations are not monolithic entities that kind of solve problems easily when they're created regulations are 10,000 page monstrosities that cause issues and create issues and solve issues all at the same time and unless you know what you're regulating I'm not entirely sure
what you're like you're creating all this work but why I mean what are you kind of solving for ultimately so the the EU regulation I find I mean I I'm in complete agreement with you of course because like if my if the point is that you do not want to encourage certain negative externalities like you do not want people to have deep fakes just to take one
example there is then you can do something targeted and specific against that when you create high liabilities on some in against somebody who created it and you know uh uh like disseminated it there's a different set of ways that you can go after it the only way to kind of go directly after the creators of AI as well as kind of regulating compute is a
fundamental misunderstanding of what you're actively regulating for in the first place so to me the question is a little bit like the people who are asking for regulation almost wanted to kind of roll back gp4 they did not want llama to to be open sourced and like we we're here eight months a year afterwards like can you admit you were wrong like you know the world hasn't
ended nobody has died like more people are killed by vending machines than gp4 so at some point like we got to kind of admit that the the the core structure by which the argument is created is wrong and there is one argument which is the only argument whereby that particular form of Regulation makes sense which is the yowan argument that like if you
don't stop it now we are dead forever if you believe that but then you have to say you believe that right then The Regulators have to say like if you do not if you let the creation of GPT 5 come forth or like llama 3 come forth forth all human life is dead that's what we're scar scared about therefore we should regulate but nobody is saying
that because that is absolutely crazy so as a consequence we're all kind of you know tiptoeing around the actual giant issue that most people don't agree with which is that you're not all going to die just because llama 3 is created it's just not going to happen people will get positive impact and people will get negative impact and there will be costs
and there will be benefits and as with all things technology so far the the benefits seem to substantially outweigh the costs and if you don't think that is going to be the case here then please please please write a paper so that I can read it because like I that's all I'm saying I'm saying like you need to demonstrate what the cost of this particular situation is so most of these
regulations are effectively trying to force uh a an a top- down view of what this ought to be like and to me it feels a little bit more like like in the the movies where the governments want to regulate a particular industry out and in some sense to me it seems like EU um UK in some sense parts of um parts of the uh the the American legislature are
all discussing it with their eyes firmly on like Mark Zuckerberg because like six seven years ago whatever everyone is really angry that nobody regulated social media or regulated them enough and now it's almost like we didn't regulate them enough and apparently it caused all sorts of problems so now we have to regulate this enough and I feel like the
I feel like this transparence of the argument is worse for both sides because even in social media like I'm not you want to regulate them but what do you want to do like what would you like you want Mar Zuckerberg to stop misinformation it's like that is not a regulatory objective like cuz you have to Define misinformation you if you say you want
Mark zeruk to not let any I don't know Russian hackers on his platform which has three billion people on it by the way then then we can have a rational discussion about whether that's feasible and like you know false positives and negatives Etc but like stop misinformation is I mean you can say that for anything right like I mean if you or I kind of went on and said the
Moon is blue is that misinformation I don't know so it's all there's a transference of guilt from the fact that nobody did anything about social media and there's a huge amount of u fur combined with the fact that there's a huge fear mongering around the fact that these things can literally kill everybody and as a result we're getting
ham-handed regulation coming top down about you need to license everything license it with what with respect to what are The Regulators coming up with the evaluation methodologies like I'm saying are they good enough at doing that they want to test the frontier models great let them do it we should also do it like is there is would you trust a government regulator to test
software for bugs before you installed it on your computer or would you rather do that yourself or do you want Norton Antivirus I mean this is not a we've had these discussions before with cyber and with the internet and with the mobile so it's like we we're retiga the same arguments in like a 10e cycle and it feels like we're stuck in Groundhog
Day I I love that answer for a variety of reasons um and what one of the challenges that just I I it just boggles my mind I always think of hitches that which can be asserted without evidence can be dismissed without evidence and you you do have this what I would call lunatic Wing basically asserting literally they're just screaming bad death is coming to the you know it
reminds me of the end is near with the guy with blackard on his yeah and and if you went up and talked to him and said why is the end near because it is because it is because God told me or whatever and and look I I am uh very much in favor of being realistic about the potential problems that will result from uh us developing this incredible
technology the challenge is we don't know what they are yet we can speculate we can uh hypothesize we can do all of those things but until we learn what they are trying to make believe that we know right now what they're going to be just seems like a very foolish way to go about things right and it's like yay we invented fire oh guess what one of the
greatest things in the world it also is one of the most destructive things in the world so did we say no more fire no we invented fire departments fire alarms fire extinguishers fire regulations fire do and and and this this iterative process as things move their way into society it that seems healthy to me right because you don't know what the problem's going to be it there's another
great quote which is no matter how smart somebody is no matter how imaginative no matter how clever you cannot ask them to make a list of things that would never occur to them of course and and and so I also want to this also um is one of the underlying reasons why I'm such a big supporter of Open Source versus closed source and you know the author
Robert Anton Wilson had a great uh piece that I'm just goingon to read it to you he says and I'm paraphrasing here he says a monopoly on the means of communication May Define a ruling Elite better than the Marxist means of production humans extend their nervous uh through through their nervous systems all channels of communications like the
written word PD radio internet search engines and of course now ai and then he says whomever controls the media that allows access to that particular thing and downloads it into the human nervous system essentially controls the world remember the old one he who controls the mediums controls the world indeed and and and so M my challenge with a closed source is
to me that just evokes the images of Pano panopticon controlled by very few people can and and it if Wilson's right by extension controlling what becomes part of every individual human's brain right right wow that's that's a lot of control and yes so you know what do you think do you think that that's an overstatement of of the power that this particular
technology could have I I don't believe I mean I don't believe so I think I think that is broadly correct I I'm there is this uh you know talking about regulation there's this like tacitus code that um uh something like the state is most corrupt when the law are most multiplied something like that yeah which yeah um that's that that's a great
that's a great quote by the way where the more corrupt the state the the the greater the number of lws and in some sense like I think about that particular quote in the way of like there's a lot of things you can litigate in little bit by little bit but overall you have to have some level of faith that the people surrounding it want to use something for
the best situation I mean that's one of the ways I think about it or like in a different put it differently I think about it as the difference between like a society led by virtue versus a society led by rules for example like if you're led by virtue you kind of know some of the right things to do and you might need bit of God rails but for everything
else if you're continuously maximizing you need morals the and and to me like when you're thinking about the regulatory moras that kind of surrounds it this is what reminds me of where you're trying to regulate it bit by bit slice by slice instead of kind of stepping back and asking the actual question what do you wanted to do in the first
place and I think like if you think about it as a fuzzy processor there's an enormous amount of benefit it can bring to all parties here because like like I am a fairly sophisticated customer government regulations are confusing as hell to me you know tax is confusing insurance is confusing and like if it's confusing to me to the vast majority it's effectively a worse Black
Box than Jud GPT is so do we want something that can relatively demystify it I think that would be useful but at the same time if you create the technology that can do that but put it entirely under is saying you know there are certain types of I don't know things that you're allowed to say or not allowed to say then by definition you're
creating either a panopticon or like a misinformation Fiat right so um I'll tell you one like one of the one things that I've been working on since like one of the I I created a bit of an evaluation suite for a bunch of llms for for for various reasons and I ran it against a bunch of the Chinese LMS um because I could I mean there's no reason
to so then the interesting things that come out from that is that they're really good first of all I should say that however like they're also clearly slanted in what they're actually allowed to say like if you ask it any questions about things around geopolitics it's like hackles get raised a little bit and like it says specific things if you ask
it questions about economics it's hackles gets raised if you ask about politics of course sometimes it just refuses to answer uh you know don't even mention Tiana man square like there's a it is fascinating to see that like it has created an actually useful tool which is like does coding really well and like you ask it to create asky art of a dinosaur it does pretty well you
ask it to name I don't know planets in reverse order with different whatever uh different languages for each like it does the things that you would want it to do but it also means like you cannot put it into production anywhere you need any of that judgment so like you cannot use it in a financial services uh institution because guess what if you're
making an investment decision you cannot be influenced by you know things that were hardcoded into to you so similarly like the way that the only way you are going to be convinced about which ones you are most happy using are by ease of use and latency like it has to be easy to use in front of you fast etc etc but also like you can trust the advice
coming from it like when I'm going if I'm thinking about investing in something I'm not going to call my friend up from Beijing to ask their opinion on a public line right because like there's a set of information that comes back is which is clearly biased I would ask somebody that I trust and that is the benefit here so like to your point about the the panopticon getting
created you know weird way I think like we are heading towards a panopticon because if you know UK is surrounded by cameras we have visual recognition that is beyond the bazu you can go into countries now with just your face as your ID especially in the Middle East it started so that's here in some sense whether we want to or not now the
question is do we just bow down and say like hey you guys take care of the models do that and like we'll just continue living our lives or do you want some semblance of the benefits of that in with you I mean I I I grew up in India right like so I had an Indian passport for a very long time which means traveling anywhere is an enormous pain in the ass and like you have to get
all sorts of visas and stamps and I wished for a big brother to exist who just knew everything about me so that I didn't have to do it like just I'm like just take everything you know what do you want bank account address take all the information just stop me from filling out forms we have the first we have the second part we do not have the
first part we have something that knows everything about me I still got to fill out forms so I'm like maybe we should even the playing field a little bit Yeah and uh I I think that the thing that scares me is that uh bit your last bit there uh I I too when watching One of the um Matt Damon um movies where he plays the the superhero um and and I'm just thinking
God it would be really great if we actually had those guys yeah you know the born identity yeah right and and and then we we don't so we have this kind of idealized version like yeah of course we can just send some of our Super Troopers over there and they'll take care of everything yeah but we don't have we don't have those guys
it's exactly like it's like big brother is actually little brother who's not so terribly bright even though he's got access to to so much vast information uh you also posit in the book that AI will be very helpful in reducing said bureaucracy that we're discussing and lamenting right now um and and one of the thoughts that I had as I was reading it and would love you
to elaborate on your view is how do you think in that specific domain use case how do you think it'll be able to navigate the principal agent problem without introducing New Dimensions of complexity or ethical dilemas Etc and it if it does that how do we solve them so I think in some sense it will do some part of that but think about the current status quod
where we already have a principal agent problem or rather a principal principal agent problem where I'm the principal who wants something done I am also the principal who is trying to do something then there is the agent who's actually going off and kind of whatever the the the the bureaucracy the interface issue that in the middle what I would what the one part
where I think it'll be super useful is purely the unstructured to structured and structured to unstructured part of it where if you imagine a large amount of your bureaucratic um face offs against the system it's because the system has a uh uh monolithic view or at least like a you know box checking view of a large number of pieces of
information that are relevant to you know your chrisan and try to get those pieces of information if it matches it it like does a check sum and your Happy Days right I mean some form of that the pain for us is like a we don't fully understand what question is being asked and B like once you understand it the amount of effort to go to to kind of get
those pieces of information and actually do it in the right order and fill it and like file it is enormous I mean I don't know how large the tax filing industry is in the United States but it is it is it is way bigger than it ought to be right like it's a it's a you know the these are examples of creeping complexity that hurts everybody on an ongoing basis not to mention things like
health insurance or any kind of insurance for that matter not to like purchasing anything stock purchases houses like there's an enormous leg level of complexity where you do not have a solid interface so given um predictability in terms of being able to convert your unstructured pieces of information and needs and wants and
hopes into some form of a structure that can interface with the bureaucracy same on the bureaucratic side for them to read this incoming information and actually clean the useful bits rather than filing it away and hoping that something beneficial will come out of it on both sides I think it can dramatically reduce bureaucracy the ethical questions come about later
once once the the they come about in the form of like number one what happens if a mistake is made and the to me is like this is an empirical question it's a little bit like are the mistakes made going to be more or less than it is made today and if it is like less you're fine like you just have to if I file something and is wrong you refile it because that's what happens
today right that's the normal thing to do and if the answer is like yes you have to refile it occasionally when things go wrong wrong however it is okay even if you have to refile it because like your AI got it wrong is an okay excuse and it's in fact a way better excuse than my CPA got it wrong which happens way more often like than than it
than it has any right to or my doctor got it wrong happens way more often than it has any right to the second part of it is like how do you trust the AI to be able to do that on your behalf and that is one where like the only way you will trust it is if you actually see it used enough times you can trust it in some sense if it does fine on several times
then you were like it kind of knows what it's doing in this sense and you kind of need to do a overall sense check perhaps or you might just use it as I use it regularly on anything regulatory or legal that like or even medical I sense check right because AI is tireless and I can give it hundreds of different scenarios and ask it the same question
in different ways it gives me a comprehensive list of questions like I did it recently uh for sort of a a family member had a had a health issue and I was like one way that I would check it is I want to get the holistic sense of information because when I talk to the doctor I get few minutes whatever right I mean you get one view however
once I get the entirety of the medical file I can actually slice it in like a hundred different ways and collect my thoughts and give a sense of these are all of the possibilities and hypothesis and use that to have an informed conversation with the doctor so then his 10 minutes is used much better by me and him because he's not giving me information I already have all the
information now we're just actually talking about questions and answers which is a much better way of interacting with it so to me like this is a the the the the way that it reduces you can see that it already it reduces complexity in the interactions that we already do with a large number of these impersonal institutions it can dramatically increase the number of ways
in which we become far better agents of ourselves like if I have a I don't know tax filing issue I don't even know where to start in terms of defining the question to ask it however if we can actually get better at that that is better for me better for the tax agent better for the IRS like everyone this is one of those few situations where
everyone's a winner perhaps except the CPA um that I used to pay a huge amount of cash to so like there's a set of different ways that you can think about where given uh you have to evaluate it you have to understand whether the answers that are coming out are sensible and straightforward that is on you we'll have public ways of doing it we already
kind of are starting but we need far better evaluation metrics in in my opinion and then having done that you can graduate the usage of those in real life situations with a little bit of the current example which is there's cavat mtor and maybe worst case you will have to create a Legal Shield around it where you can purchase insurance against it so we can get there even with the
current AI which surprised me honestly because I wasn't sure that that was the conclusion I was shooting for it'll keep getting better but there's a number of ways in which we can finally for the first time at least in my lifetime got the red tape involved in actually interfacing with imp personal institutions yeah and uh you you made a point there that that I uh would would
love to extend a little and and that is the compared to what you know it's often the case that uh you know as as you point out well it it got it wrong so therefore we will never use this particular technology much like they do with driverless cars right a driverless car a driverless car crashed and killed somebody take them all off the roads
they are absolutely not to be allowed well wait a minute compared to what you know how many crashes killed people with a human behind the wheel a hell of a lot more than with the driver list car so it seems to me that we set we have these planted axioms in the argument which sometimes make me wonder whether the other side is arguing
in good faith because like compared to what's actually happening right now when people are filling out their taxes or getting their medical information or getting their insurance or compared to some state of perfection that will never be reached by anything never has been reached by anything ever in the past and and it and sometimes it just drives me
mad when I see these comparisons like and compared to Perfection well yeah then let's just not do anything let's all sit in caves and and just you know be nothing I precisely right I think to me that you know it exemplifies the the the I mean people are always worried about what things can go go wrong because it is much easier to see than
like how things can go right because if especially because if you already knew all of the ways in which things could go right then like you know we would be living in a very like weird world because you can't predict the future I like when um it's like when the first iPhone came out like everyone was predicting all the ways in which in which it was wrong right like right only
had 2G didn't have 3G couldn't do copy paste was heavy like glass would break if you if it fell and guess what all of those things were correct cuz like none of it mattered because like we moved past it by teaching us and teaching them how to actually use this new thing in the first place the same way when computers were started getting
introduced into workplaces tons of complaint you know it'll get things wrong and like you know people whatever all of the complaints were correct however we changed around it which is how we do things better we do things like far more efficiently which is the kind of metric that we Ed we should judge these things by so I feel like a lot of the times when you argue against
some sorts of progress it is much easier to hold on to like oh yeah yeah but see it got it wrong here which is the thing is like the question to ask is can we fix it and also like is that error okay to have now there are some errors that are not okay to have you don't want to have a I don't know an accidental accidental nuclear explosion right fine
like so there are certain systems where you do need high levels of determinacy and incredible numbers of redundant layers so that some accidents are just too bad and we do not want it to happen but that's not what we're talking about we talking about helping people's lives as it exists today with enormous degrees of error I mean the number of people
that I have spoken to who have had uh medical misdiagnosis legal bill issues because like they were advised wrongly tax issues it's staggering I mean you asked three experts you get three different opinions and if that's the situation as it exists today then let's make it efficient and actually like cut it out and then we can come back and fix stuff
later but we do this this is a waste otherwise yeah uh again Amen to that but also as I was listening to you uh describe uh your family members Medical issue I hope I hope that it was resolved um it it did make me something I've been giving a lot of thought to and and uh concerns me a bit is that have we come up with a technology that is I don't know people who are very clever or very
smart are already at least among my group are already using uh multimodal AI for a variety of things do do you think that do you think that um and it also guides my idea like AI everywhere I when when AI becomes as invisible as electricity then maybe this worry about the super smart people over here using it to its maximum effected Effectiveness are just going to
bury uh you know the the people who aren't using it it's like like uh ahmmad Mustique who is the founder of stability AI which I'm an investor in he says AI is not going to destroy Humanity or human jobs humans using AI will destroy humans who refuse to use it uh what do you think I mean I think that's the positive view of the world right like
humans using better technology enable things to get done much better enabling all of us to have far better standards of living as time goes on that's the dream I mean there's a reason we only employ a minuscule fraction of the people who used to work in the agriculture sector compared to like you know the vast majority a century ago like that that is progress
so the question now is twofold right I think you know there are two problems problem areas problem number one is like to our earlier conversation what happens if the mistakes that are being made different mistakes to what the people are making which will happen but like that's what people are scared of not that like it'll make mistakes but it
makes different mistakes and we're used to these mistakes but like and some of that I think is cultural because it's like get used to new mistakes that's you know that is the job of growing up right in some sense so that's fine um the second bit is like what happens if this happens really fast which is a worry and a bigger concern and potential social
issue potential I say and in the even if we pause it that is likely to happen even then like stopping it from happening by Fiat is not going to work so to me like the question then becomes we should perhaps much push the question of how much better how much more you need to experiment and innovate in your daily work so that you can actually do
more things far more rather than saying like let's go the Lite way a little bit and kind of like stem back the the speed with which the machines are actually coming in to kind of take our jobs the good news here is that the amount of effort it takes for getting an AI to take your job today is remarkably high it's not small except for like I said sort of small bits and
pieces where you can actually highly and tightly Define the job scope and therefore make it amenable to automation or Creation in some sense but that was kind of always the case right and I feel really really bad for people whose livelihoods are being impacted by doing this and we need to do everything in our power to help them socially as well as
educationally in order to kind of do better but stopping it is not the right method to do that like um it's like it reminds me a little bit like the closest analogy and I was thinking about this the closest analogy was like uh like stocks used to be traded by hand right open outcry it's not anymore and this didn't happen over Generations it happened
pretty fast like you like Electronic Training came in and it kind of moved things about and like guess what people found new instruments people found new ways to trade and like a whole series of people needed to find new ways to actually deal with the reality and which was much better and far more efficient for everybody concerned and like the fear of um potential displacement or
like the slight level of complexity and Chaos should not stop us from saying like let's not bring in the benefits which are you know 100x THX what the costs are because we can see the costs like we can actually see it and I mean this is an election year every year here at least in the United States you talk about this like whatever Co belt and so
on about you know boding the loss of certain types jobs and I think part of the reason that exists is because it like you need to have the conversation that like there are certain jobs that once gone don't come back for good reason because they're not as valuable anymore so we need to get better jobs which is the right way to kind of go about it now like there's that
adjustment period like you need better you know social safety nets and a set of ways in which you can help the people who have lost the jobs but there is no way to kind of stem that tide and I don't think we should try to because I feel that would be doing a disservice to humanity can not agree more with you on that it's like I was chatting with a
team member yesterday and talking about the ice harvesting industry in New England yes right so thousands of men lost their jobs when we electrified kitchens they used to literally saw the ice from the Frozen River Transport it to the uh processing uh facility where they would cut it into chunks with sawdust on it Y and then the icean would take his truck around delivering
ice for your ice box which some people still call refrigerators right but there's cultural continuity there but those were bad hard jobs yes I mean that's the thing the same thing exists in logging right I mean like it's it's there are so the history of our I don't know species like I had this deep dive into progress at some point and one of
the cool things is like the number of ways in which we have bettered ourselves by leap frogging is phenomenal which is the most beautiful part of it like name anything wailing another uh not to mention exceptionally hard and dangerous job that nobody actually needs to do anymore um but I think the speed of adjustment here is one cause that often
people kind talk about but the solution to the speed of adjustment is adjustment on the other side rather than you know uh naively stop everything in the first place because even if you na even if all of AI progress today just stops like nobody does any research anymore the progress the speed with which it actually implement the real world is not
going to change all that much because it's just there's there's way too much benefit in these Technologies to not be able to use them in the first place so it behooves us to know how to use it completely agree and uh you brought up creativity which is another thing you cover in your book uh and one of the questions that I had because I share your view I uh I share the view that
this is going to be almost like Godlike powers for humans uh and leverage for their for their creativity but you know one of the things that I've thought about a lot is essentially you know I guess we could uh uh borrow a title and call it a critique of pure reason um a lot of a lot of uh you know Peter watt says we're not thinking machines we're feeling machines
who happen to think and a lot of creativity uh comes from these leaps of imagination um where I mean I think of Einstein I think of Tesla I think of you know a variety of people when you look into when when they were talking about how they came up with ideas uh it was not in a rational or any way order way it was literally a leap up
the chain so so to speak uh and many would say that was an intuitive leap up which leads me to say that I think that much of the use cases around creativity are going to almost require a Sedar model I.E man plus machine show me where I'm wrong help me out uh where where that's an incorrect view I I don't think you are wrong I think the the the only
caveat or perhaps addition that I would make is like centur models work best in areas which are not directly um entirely competitive with the same things that the AI do unless you find joy in doing it because then it's a self-fulfilling kind of Prophecy so to me like currently and at least for the mediate future AI is best used in areas where you can
either automate part of your own job and yourself and also use it together with you in order to make your ultimate goal better it's just like any Tech we are all centors already right like we live most of our lives on digital technology connected with other human beings we are part of some weird form of a hi mind and we are all kind of cyborgs this is a
fact so then the question is how much more integration would you like in different facets so that you can actually perform some of these things better and the answer is all of them now there might be some things where guess what like if you like drawing for fun you're probably still going to like drawing for fun despite the fact that if
you do want to kind of make a profession out of it there are some things that the AI will be able to do much better and you as somebody who actually understands it and can use it better and knows the intricacies of drawing will be able to direct it and make use of it in ways that like me as somebody who doesn't can't like you know your knowledge and
education in doing that particular thing translates to how much better you can actually do something it's like like giving yourself a boost right everyone gets a boost kind of question the that's that's kind of not going anywhere so I feel the to me the the education part is super interesting and important because ultimately the only way we are going to
be able to make our peace with it is probably the wrong way to kind of put it but at least like the only way we're going to be able to kind of utilize it to the largest extent is if we get much more comfortable using it on a regular basis um and occasionally I've had discussions with like artists and people who kind of feel very strongly against um Ai and I respect
some of their opinions not all of them at all but like around the fact it does not have a soul or you know those kinds of opinions and I'm like leaving aside all of the metaphysical stuff I get what you're saying but also that's your job right like that's you know what I mean like why would you expect it to have a soul like your work with it is what will
imbue it with sort of the creative characteristics that makes it better like that's that is the entire purpose so which comes back to your centor model but what that means is like I don't know you just want a silly illustration to do something for your blog or like you know your little product launch you don't care like you don't need it however what
that all if you want to do anything genuinely creative and artistic you need it because that's the only way that you're going to be able to um create something that is state-of-the-art because we're we're all raising the waterline right like I mean it's a um it's a continuous pushing up of what is the best that something can be doesn't
mean that you always need to beat it but at the very least it means that you have your aspirations are much higher than where they used to be before and you do kind of need to like fight for that I think the uh it's the hline Cote you know like man is not a rational animal he's a rationalization animal like it exists for a large number of these
decisions we are not thinking about it rationally we are rationalizing by saying I hate the fact that a machine did it therefore I think it is bad as opposed to going at it the other way around and solving for the problem that you want which is I want to create something a a novel a movie an image anything and I want to use whatever tools I have at my disposal to make my
vision come to fruition as opposed to sort of flipping around and saying oh no none of my Visions have anywhere to go anymore like we are all going to be stuck here forever which is a defeat toci to do I don't think it's helpful yeah and I that's part of human OS I mean there's the probably an apocryphal story but the first uh Sumerian CIA form found was a lament
that all the good poems had already been written um but true turns out that was incorrect uh but uh as far as the the artists as as we were chatting before we started to record I I've been collecting C art for a long time I have a lot of friends who are artists uh many of them have come to me and since we're friends they're like hey I'm feeling like really exposed here
yeah and and I I give similar advice to what you just gave we actually invested in a uh seed Stage Company called Juan and what it does is it's actually uh a device with with which you can draw but it's AI and what it allows you to do is iterate against your own work and we have been having that once it gets in the hands of sometimes even formly hostile artists
yes it's like night and day they they use it for a day and they're like oh this is the coolest thing in the world because essentially they're their understanding I the way I look at it is they're not thinking about it in terms of a I they're thinking about it in terms of oh just another tool and oh my God look at what this tool can do it can
iterate yeah you know they do the sketch especially with graphic artists love it they do the the sketch and then they say to the AI do 10 variations on this and the AI does and they're iterating on their own work but it's like having like the best sparring partner in the world yourself you know what the cool thing about that is that there is an implicit
change there where before you start using it you kind of confus in your mind a bunch of Notions about like sensient and Consciousness and a alien being and all of these kinds of things associated with something that produces uh subar work better work whatever but once you start using it none of that matters it's a tool like exactly you suddenly the the
the boundaries become slightly less fuzzy you realize that it doesn't have wants and needs and whatever but you can use it to create what you want as opposed to like it existing as an independent entity with you know whatever alien senss that it has on its own so the the funniest thing about the the the the funniest thing to me in this entire process has been like you see
that there's been an enormous number of questions and decisions that have existed in uh you know neuroscience and Consciousness research with uh depending on who you ask minimal progress um and in many ways we now have something that we can actually poke at the neurons right like the the the the highest resolution image of the human brain
fails in comparison with you being able to poke at the gradient of like a you know um attention based Transformer model because it's just now you can actually test some of these things and actually have early stage very minimal empirical some understanding of what might be going on in our heads it doesn't translate one to one we are not Transformer models internally not even
the individual cortical columns I suspect however that is we finally can do Empirical research in some of these problems which we have all put in like a it's like God or you know infinite you you put a placeholder around it there's a bunch of feelings associated with that word and anytime that word gets triggered all of those feelings kind of
pop up as opposed to them being active you know like like a C++ pointer kind of pointing to something genuine this just ends up being like this amorphous blob that you use only when you want to get angry at an argument right but you know on that point like if you consider integrated information theory for example which says that Consciousness develops from
the integration of information in complex systems with a through a variety of uh of that happening you also make the argument in your book that uh AGI doesn't need to be conscious um what what about it if we look at it through the lens of integrated information Theory and it Consciousness spontaneously emerges do do you have a plan for
that uh I don't have a plan for that I hope for that but I don't have a plan for that I think I have a I I have a friend who's a wag and he said just pull out the plug I mean honestly yeah but like let me put it this way right so they um one of the clear thoughts in my mind when I was writing the book was that I wanted to be able to explicitly
explain where we were without needing to pause it um huge new future developments that might come so which I really appreciated about the book by the way yeah because like I feel like once you make once you expect the future development to do X then you can kind of justify anything you want right it's a little bit like uh uh you know whatever James Cameron making his Avatar movies
every every like he just kind of expects something crazy to come about and then like Waits 10 years and something completely different pops out but like um so with respect to IIT I'm I'm not an expert although like I I read a little bit about it I think the the core conceit as I understand it or at least one part of the core conceit is that it
it wears fairly close to like a pan psychism argument where there is sufficient levels of information that is integrated is what kind of pops out a form of um sension or Consciousness the problem to me is that I don't see how that's falsifiable in in any meaningful sense and as a consequence I am perfectly happy believing it in the abstract however I
don't fully understand how to think about it in a real world scenario that said let me kind of place it in context we we are mostly talking today about like you know large language models to a large extent I mean we briefly talked about stability and like the diffusion models and stuff like that um all of these are even though they're
powerful highly blunt instruments right like I mean you have to throw in terabytes of text in order to make it create some interesting semblance like I mean I did this weird calculation at some point and I think it's like whatever maybe 10 to 50x as good as um the current uh zip or whatever other compression algorithm so they are much better at retaining information and
doing cross connections between itself now you know there was this Yan L's like recent video somewhere where he talks about the difference between you know one of these models and like what of four-year-old child sees like the amount of information that that that here she sees by the time she's four and it's like a huge amount of information that
they kind of taken in video but the thing is like for us as humans or any sensient animal that we know of they have a a a pre-made template uh helpfully Guided by Evolution inside our heads that enables the processing of the information in the right format but they also have an enormous multimodal set of receptacles that can take in information and actually integrate it in our mind
they might not be all Transformers or all diffusion or some new architecture thereof it's a some way of making sense of the fact that like you know a 2-year-old walking around sees something uh says something hear something and is able to integrate all of that together and then be able to come up with some sort of a world model that he or she
accurately perceives right there is a chair there is not a theoretical sentiment or a hallucination but it is something that they can accurately perceive and that level of groundedness is something that no AI today comes even close to actually having which is interesting because what that tells us is like there is a large number of leaps that we might need to make in order to
create something like that so to me the question is um I defined it in the book as something like you know if you want to do a scientific experiment which involves some level of supposition and some levels of like induction how could you think about creating an AI that is able to do something equivalent and the answer in some sense is like we probably
don't need to think about it in terms of completely new architecture but can we try to get there through a fair bit of hard work from what we have today by actually connecting them together creating an infrastructure around them having the right kind of redundancy having the right kind of memory etc etc etc that is the kind of Crux of the
question because when we kind of posit it closer towards what we are which are much more complex beings that run on whatever 20 watts of power that is a um order of magnitude different and at least my hypothesis is that we're not going to stumble upon Consciousness by mistake I don't think there is a um large enough Corpus where automatically
it just starts up because a autor regressive models don't work like that I don't think that's how Consciousness emerges in order for that to happen we would probably require a Transformer equivalent llm for each cortical column that is actually interconnected and trained together with multimodal information that is coming across sort of I don't know call it years but for
them in like you know subjective time maybe not years including things like proprioception walking around like like there's a set of modalities that are required for that innate sense of the physical world to be embedded so that on top of it you can actually learn different things and maybe at that point would there be senss that emerges or
Consciousness I have no idea I mean I hope so because I'd love to meet a alien life form in my life time I'm not nearly as scared of that but I'm not entirely sure that that is a I don't think we should make decisions on the basis of that because nobody knows at this point right I I I I agree let's let's shift gears a bit because uh you spent a lot
of time as a venture capitalist uh and still and still are are are doing that and and I I I'm curious about your your view on what's going on with the the the tech Giants we all know who they are Microsoft Google gole Amazon Nvidia at all do you think that they're creating potentially hostile environment by by doing uh a little bit
of a shell gate and and saying that you know we're investing 10 billion and then you investigate fine print and nine billion of it is credits that are going to go right back to their cloud computing um and and and that they're doing it at these skyhive evaluations that seems to me to put the you know good old invented in the garage startup in almost an impossible
place what are your thoughts on that that uh disconnect between that sort of funding which is going to very specific players like open AI anthropic Etc and and and creating this hurdle rate you know I remember what is open ai's last uh valuation like 89 I think or yeah n90 billion doar right and and you know especially when you have the
slight of hand with most of that being pedits that go right back to Microsoft yeah um and then and then you've got to create enough liquidity in a secondary market for all those open AI option holders who are like oh I love90 billion dollar valuations it just seems to me that it's a it's it's not an outright flat market failure but it does seem to be tilting the playing field
quite a bit to me what do you think about that so um as someone who has a bit of my retirement savings invested in the big Tech I I you know I would say thank you SAA uh he's done he's done he's done he's done well historically like if you think about let's take a couple of examples in Tech right IBM M used to do this thing um somewhere in the mid 20th
century later half of the 20th century where computers used to be incredibly expensive in the main frame era and they used to help companies Finance the purchase of the companies um uh purchase U main frames to me it reminds me a little bit of that because like because IBM had so much money they could fund their customers to become their customers and get like a little bit on
top effectively and it reminds me a little bit of that right I think did rollro do something similar I I have this vague memory that rollro did something similar as well where they like um I yeah that one IBM story is well known but I I don't know the rro I think they did something similar with u uh with Airlines where because like they want to do uh paper
use or paper hour or something like that instead of actually paying it up front because Airlines couldn't do it and therefore they got a little bit of bump Etc but anyway um historical asides aside I think the interesting thing here is that um the big Tech dominance in uh Tech is clear and it's clear in sort of a couple of different ways if you think
about the largest tech companies today most of them are the largest tech companies 10 years ago like if if you think about the largest Venture back tech company since I don't know that ipoed since 2012 or something it's probably like since Facebook effectively there's not been one that has crossed billion I think like service now is probably the thest that has come to it
so in some sense um the only arguments of companies which are not against that in the private Market is stripe which hit close to 100 then came back down uh space SpaceX which is Elon magic so every number is imaginary over there and I don't really know what the right valuation is likely to be but there is a dirt of enormously large companies that
actually exist in the Venture pack Market which used to exist in the prior market so in a weird way the way the VC actually made money is not just by kind of making sure that the companies become as big as Google but actually acting as late stage financiers so that the public market effectively got privatized so like you would have been able to
purchase some of these public companies at like three four five 10 billion valuations worst case scenario I don't know what Facebook been public at but it wasn't anything crazy uh compared to it was crazy then not compared to today but instead you would hold on till it's it's like 100 billion and it goes public and it kind of Treads water or like goes up
a little bit comes down a little bit right like snowflake is a good example what this basically means is that for today's AI companies has kind of caught in a bit of a bind because on the one hand they have enormous capex requirements if you're building foundational models or gpus Etc the hyperscalers have more money than God and they can kind of as they can
actually pay them to use their credits which means the money they're giving out mostly is coming back in as Revenue if it comes back in as Revenue to Microsoft and Azure that give them like a 5 10x boost instantly which in their in their market cap so they're like this is a perfect trade for them on top of the equity and the benefits that they're
actually getting what this tells me is like that game is kind of shut for startups to play in if your game is I'm also going to get some gpus and I'm also going to play in the hyperscaler game and I'm going to provide access to I don't know fine-tuning models whatever it might be I don't see a pathway to creating a large company basically by doing that I
genuinely it reminds me a little bit about elastic search fight with u AWS and because it was open source AWS roll there on and they're kind of like it's history at this point right or many of the database fights at this point so if you're a early stage like startup to me there are two paths which are kind of not TR upon Tren upon by any of these
large companies uh which is number one go vertical because like no matter what you think about how smart these models are how smart they'll get in the next year or the year after that the world as I have seen is infinitely complex it's exctly complex is a better word which means like if you wanted to do a job in a particular domain the only way for you
to do it is to explain that domain to the model and help understand what it can actually do and not do this is not trivial this is in every domain is complicated um I was talking to sort of some Advanced manufacturing folks recently like if you think about 3D printing it's heavily complicated 3D printing Metals is complicated at a molecular level because you can't
predict exactly how the molecules react at any point the powder has to be in the perfect format this is not something an AI can just like figure out just because it's infinitely smart so as a consequence the deeper you go the more narrower you go the better chance you have at creating an actual company that builds interesting and useful things
number two is that you can try to shoot for genuine algorithmic um I don't know step changes if you will you know Transformers came out from a bunch of researches in Google in 2017 could you create recreate something like that quite possibly I know folks like Sakana are trying to do that in Japan you know so there is a possibility for some new revolutions
there it's not really venture capital or maybe it was it is Venture Capital as it was practiced in the 80s because effectively you're funding a research lab and hoping that research comes out so it's closer to like a privatize Bell Labs that you're funding and maybe it works and maybe it doesn't but it's a it's a weird new form of funding Dynamic
that nobody's going to call on to so put those two things together what do I see like number one I mean IBM used to do this they had the heavy Monopoly and the only reason they got sort of kicked to the curb a little bit was a antitrust made them incredibly paranoid about doing anything lawyers basically ran the company for 10 years and B like there were genuine new innovations that two
people in a garage actually could do which was kind of a algorithmic stepup to create some kind of software that could actually take make an inroad into this um this particular domain so I think it is something similar I do think that if you think about the best startups to come out of this era the chances are I don't know whether the
chances are that all of them will require $10 billion of GPU to be successful because I think that's not likely to happen because if it required capex investment then there are already enough people who can deploy capex right even if you even if it needed a a billion in capex investment the number of companies and organizations and institutions and individuals in the
world who could deploy a billion into one of these companies is staggeringly large so it cannot be capex it has to be some combination of unique inside and knowledge and um uh knowhow that needs to happen for you to be able to create something that is Meaningful to a particular section so that is very much where my kind of lens is with respect to
the market that you have to solve like they the big companies have built the base very nicely and the base keeps going up which is wonderful but if you want to create something you have to solve someone's real problem on the ground and all of the things that we've been talking about during the conversation today whether it's on a getting better evaluations understanding
the process better if you're going to be a you know you need to be embedded in those individual domains for you to be able to do it so I don't know what another analogy is like Cloud when it came up there were a bunch of other Cloud companies that tried to start and compete with you know AWS and Microsoft and Google none of them really succeeded
right like even the even the big tech companies didn't really succeed like Oracle has its own cloud but it's not like super successful the only ones that did succeed was saying like oh you're providing the base we will actually go ahead and make that useful and make use of that to build something genuinely meaningful on top of it um by providing
something that you couldn't get from the regular Cloud companies whether that is you know folks like snowflake or whatever data braks name any company that is going to come out of that era you have to provide a service that is Delta on top which two people in a garage can genuinely do because it is something overlooked in correctly by the
large companies in a particular Niche where they genuinely just don't see a huge amount of value yeah um I C I could talk to you all day I'm getting the hook from my producer here I always babble on as he says but I got to ask one one final question and then we're going to get to make you the emperor of the world again um okay so the question revolves
around something that I've been hearing more and more about and from really really smart people way smarter than me and and it's the idea that are we running out of useful data to train these models on I have an opion about this but I want to I want to hear your opinion um and and then wrapped in that a question do you think like uh the some of the largest purely datadriven
companies wouldn't it would it maybe be smart for them to acquire a fully integrated uh AI outfit um because that data is behind their firewalls and it isn't excess it isn't on the pile so to speak right which is what all of these other ones were were based on so to so two-part question are we running out of data and I'd also like your view on synthetic
data I also have a view on that but I want to hear yours and then do you think that might be clever for some of the bigger data companies to just go acquire uh a full stack AI company married there behind the firewall data uh and yeah God knows what happens I'll do the second one first the answer to the second one is yes if they can afford
it I don't know whether they can honestly but but yes now now comes the caveat which is actually the first question so there was this big like especially over the last decade everyone started talking about data as a new oil and we finally found a place where data is the new oil but there is a problem here which is is that while we need large quantities high
heterogenity and high quality of the data that is coming in a priori it is very difficult to use that data to help easily extrapolate in some sense to what the benefits of using that data as likely to be so in some sense um you know the perfect example is maybe something like Bloomberg GPT which trained on proprietary data that is very hard to get for very large sort of you
know many people turns out like they didn't get a huge boost by just using their data because like which is uh one case in point and there are others of how it is insufficient for you to have data that nobody else has for you to be able to make use of it and create your own model instead what you need is data that nobody else can have or extrapolate
towards which is a much higher Bond right because the lln are in some sense extrap ation engines it finds implicit associations within all sorts of information that is fed into it so implicit in that is the assumption that if the data that you have could have been extrapolated from existing data in some sense which a lot of the data behind firewalls are then effectively by
itself that is insufficient that said I do not believe that we are anywhere close to sort of the end of data in some sense because there exist an enormous number of new modalities and capture mechanisms to get data that we don't have I mean we're kind of today at like the uh p++ era where we're talking about internet Corpus or taking some non-
internet Corpus and making it internet Corpus but if you think about the last vast majority of data that exists around us it's not just the data that exists in the internet right it's multimodal information that comes from everywhere like it is the um like proper reception models it's a like there's a set of types of different types of data that
you cannot easily extrapolate from existing data sets which is kind of where the value is so um that is my broad assumption what what does that mean practically sets of data that I'm super interested in um interesting uh visual data I think is uh vastly underrated audio vastly underrated um name your modality smell I don't know vastly underrated uh things like
there was the Berkeley researchers who helped the robot walk around in order to capture data the data about how it works and how it learns in itself is actually vastly underrated and lastly specifically in terms of companies and organizations process data vastly underrated like and illegible mostly it's not captured but that's where all
of the value is right if you think about it you think about how do you make decision inside a large corporation or a company like yeah you have some written documentation but that's not what matters you know like I talk to them they talk to them we do a meeting together we capture a bunch of information we write it down we iterate on it then we talk to three other people
do some soft like it is these 200 different things that interact with each other that actually creates the final output and that process has almost all of the magic instead of the end point so I think there is an enormous amount of data that we're not capturing synthetic data is interesting one I am relatively positive on it as um I I this mind small
thing to test like a hypothesis that could you tune a GPT on its own output like a long time ago and it was it was cute and it was cool and it worked and um I was like curiosity satisfied but recently there was a new paper out where people actually took a new method of training DPO and applied it to the outputs and used the same llm as a judge
and suddenly started performing much better now the caveat here is that I do think that this will work much better in in terms of training the AIS to be do certain things much faster or it's like the Coan Paradigm Shift versus in Paradigm progress I think synthetic data will help dramatically in in Paradigm progress I don't see how it can help it
shift in the in any meaningful paradigm shift because it has no way to extrapolate the information that it doesn't have but within the information that it has it can create much better data and be able to jump up um in the so it can go from I don't know uh llama 2 to llama 3 Performance by using synthetic data and all sorts of clever tricks perhaps but it's not going to be
able to go from like I don't know gpt2 to gbd4 type kind of leaps if that makes sense it does well as I said I always have a wonderful time talking to you rth um now we're gonna make you emperor of the world and you played this game with us in the past uh you can't you you you can't kill anyone and you can't put anyone in a re-education cap and you
can't stick anyone in a strange Loop that is uh just endlessly uh has no exit but you can incept that you can incept the entire population of the world we're going to hand you a magical microphone and you could speak two things into it and all eight billion plus or minus 100 million people on the planet will wake up whenever their tomorrow is and
they're going to think I just had two the most wonderful thoughts and here's the key and I'm going to begin acting on them immediately what two things you going incept oh gosh I'm I'm really hopeful I'm not going to repeat my old ones but it's compounded by the problem that I don't remember what I said before exactly but I'll tell you what's on my
mind right now which is the right way to do this number one I think I would tell more people to read more fiction um it's compounded but like I've been having a you know over the last year I think one of the things that I noticed while I'm writing as well as talking to a large number of people is that there is an enormous number of folks who are
super excited about reading non-fiction and there's a lot of good books that are coming out and it is wonderful in sort of all sorts of ways I don't I don't need to tell you that however I feel like the role of imagination and Whimsy are not g being given enough shrift if it will like they're not being given enough he in daily life I think you know not just because of AI
or anything I I think it is the most wonderful gift we have it's what it means to be human I think we're doing ourselves a disservice by not being of attention I 100% think that like that is one of the one of the things that I would say would be please do that and number two and this sounds like kind of thing that I might have said before is
I'd encourage people to Tinker more to produce more um and I mean I've said this before I think you know as including connecting to the last question that you asked me there's a large amount of learning that is impossible to get by reading almost all of the useful learning in my opinion which is I still don't know exactly how or why but there is a uh a a a certain
process where doing something actually teaches you a lot more than reading something or listening to something and as a consequence if I can push people more towards acting on any it doesn't even matter what it is just doing something I think it would dramatically improve the way that we actually perceive the world and are able to make
use of it I would love to see that happen I I love both of them my my friend David ha who is brilliant at Ai and everything else when I was chatting with him I said like what's the one quality you really like to see in someone you might want to hire and he said they have to love tinkering and and and so I just uh I I also I think about that a lot and I and
and I and I love both of those obviously well thank you so much for rejoining us I'm sure we'll have you on again just briefly uh tell people where they can find you online I am on uh strange canon.com and I am occasionally on Twitter um you can find me at Christian RIT perfect thank you so much my friend until next time my absolute pleasure Jim I love doing it as always