You're listening to leaders of AI the podcast that cuts the noise to bring you real world insights right now applications and future shaping breakthroughs in artificial intelligence here are your host jul McCoy and Dave Shapiro when the technology gets good enough cheap and up on fast enough that's when Mass disruption happens what are the constraints and capabilities that you're really seeing right now that That people are not aware enough of the capabilities are just taking these massive leaps forward where narrow AI is just excelling humans in every single area that I can see I find
that the AI is training me as much as it has already been trained AI can already make better decisions Than People eventually AI will run the government organizations are slow Dum AI right now probably what a thousand people in the world are using AI to its fullest extent I'm very Transparent that I use AI in everything service oriented which is basically the West are in trouble that's why we picked the name first movers cuz that's who we are and that's who we serve and our mission is get this stuff deployed safely and effectively into society
as much as possible welcome back to leaders of AI podcast we have one of the most exciting guests I think we've ever had on our show we've got Iman mustto here who is The founder and former CEO of stability AI which is the company behind the open source text to image generator stable diffusion and as of January 2025 that is now worth 1 billion dollars so it's amazing that imod led so many initiatives in this space where we have so many models now so much Innovation and of course last March 2024 you resigned as a
CEO to pursue decentralized AI initiatives I've heard some of the interviews you've done with Peter diamandis and it's amazing your goal to really bring about I think one of the biggest initiatives we ever could in this year and with what's coming where we can create this new economy of a decentralized future so first of all welcome to the show I'm so glad you're here thanks for having me it's a pleasure amazing so we want to start with we've got kind of a two-fold conversation here because you're gonna be such an amazing guest and we want
to Pick your brain properly um so I want to start we're going to link to this in the show notes with your article posted on X called when Capital no longer needs labor how does labor gain capital and like I said earlier I think there's no bigger question we could ever ask we have to be thinking of this new world that's rapidly coming our way when automation takes over all human labor so how do we get there that's such a big question so one of the pieces you wrote About is of course how AI is
disrupted ing this traditional Synergy between labor and capital so question for you to share with our listeners how do you see this playing out and what do you see as far as the timeline do you see like the end of all jobs the end of this year or in the next half decade what do you see with the timeline timelines are really difficult like every time you step into AI some new Leap Forward has happened and I think the key thing is it's like a Turkey at Thanksgiving you know it's fine for a while and
all of a sudden it isn't it's like stairs up escalator down um when the technology gets good enough cheap up and fast enough that's when Mass disruption happens and so we're seeing it already in areas like why would you need a Filipino call center worker which is one of the main parts of that economy right and AI can do it better at a fraction of the price and we're not even sure what fraction of a Price like in the coming years how much does a Hollywood movie take to make or how much does your taxes
cost to do with an 01 type model so what we're seeing right now is that the models got good enough but they didn't have recall they didn't have memory they didn't have continuous learning and now they're moving from Models which are these massive translation engines that words go in and images or code or something come out to Systems that can learn adapt and improve and they don't make mistakes a lot of the discussion has been around creative AI like but how many chefs do you really need to come up with brand new recipes you know
like this whole AGI thing what really the impact is going to be is really great cooks that can follow recipes that can answer things for call centers that can do your taxes that can give you legal opinions that can diagnose your medication and other Things like that and I think that pretty much this year is the year when all of those get good enough cheap enough and fast enough across probably a third of all tasks certainly those that can be done remotely because now the eye can also see your screen the diffusion of that technology
though may take longer but this is definitely like massive economic disruption happening at scale over the next should we say five years probably Max 10 the example I like to Give of that as well is Teachers those of us that have kids or nieces or nephews every single head teacher in the world in December of 23 had to ask a question what's my generative AI strategy because of chat GPT have we ever seen that when every single person in One industry has to ask the same question no and now with some of the lat results
it's clear that those that won't Embrace this technology the kids will fall behind those that do and so you Repeat that industry by industry knowledge process by knowledge process and it happens at scale because of how ridiculously cheap it is well said one of the other items you discussed was comparative advantage which I thought was really good you talked about how it's shifting from labor intensive sectors to those defined by computational infrastructure really well said so which countries or regions do you see best position to thrive in that Switch I think this depends on how
we execute on it but I think it's clear to me that service oriented economies which is basically the West are in trouble because you're Reliant upon almost that lift basis of you know the whole Henry Ford I want to have my workers be able to afford the cars why do you need to hire any graduates you know whereas the other side more AG Aran or industrial based economies have a comparative advantage there because like when you Look at the average IQ of the world it's about 90 and it's mostly an infrastructure and energy equation thing
like most of Africa doesn't have the infrastructure and energy to give the schooling to allow people to do IQ tests it's not a base intelligence thing but now what happens if for a cost of a few dollars a month which is I don't know now like a million tokens or a million words with the new deep-seat model you can give every kid in Africa an AI that Has 120 130 IQ they have comparative advantage you know because they can work for almost nothing you are taking the brain drain and you're almost reversing it in that
all the brain drain meant that the best smartest people from all these countries came over to the west to Silicon Valley now we're packaging up their brains in these train models and we're sending the brains back so I think that those countries might have a relative Comparative advantage whereas we will face I think massively deflationary pressures within the west and a question of what do we do with all these kids graduates and then wrote workers what do we do with all our knowledge cooks that so well said that's a great question because you hit on
something that's so true it's you know we've had the service-based economy the gig economy here in the west and so what will we do in that shift what would be Some advice for you know I get a lot of YouTube comments from students that are what should I even study with this coming my way well the recommendation I've given everyone like at work and other is that like everyone takes co-pilot courses now sorry cursor courses and then they're like wait we can build apps like again the comparative advantage will be most people still don't know
about chat GPT like it has what maybe 300 million users The next most used Claude is like 10 times or even 90 like more than 10 times less if you look at the data versus 7 billion 8 billion people in the world like it's a bit crazy like every American should be using this but they're not and certainly not using it properly like again how many know that you can build an app from nothing that gives you a massive comparative advantage because it's not like this will change overnight you know like but When you look
at the timelines we're talking about five 10 years that is overnight effectively in these terms I mean can you believe that five years ago was covid starting it doesn't feel like that does it wow and the whole world changed within a year or two this time it will be the whole year world changing a little bit longer but you have that comparative advantage if you utilize a technology Now and you're like what can I do with this technology that's great that's a so um aligned with our ethos at first movers which Dave is a big
piece and advisor in and I started it to help companies literally become first movers because if you start learning and playing with this now the benefits like you said five years feels overnight so I'd love to switch over to decentralization which is something you talked about in the Article I thought some really great points Universal basic AI was something you talked about you even discussed potential Solutions like Universal basic Capital data ownership rights big fan of literally everything you laid out so what governance structures do you think are necessary to prevent something that a lot
of people are concerned about which is the centralization of AI power in too much of one location it's a hard one that one I mean the the key thing Here is that these models I had a previous piece how to think about a I'm like it's like a new AI Atlantis that we found with billions of graduates and this is leaving aside the robots which we can come to right the physically embodied AI which is also a big issue what is the curriculum that teaches them because if you teach the kids wrong they'll come out
wrong you know and so a lot of alignment is actually about the data inside models Because we're going to Outsource more and more of our cognitive capital and capability to these models it will be medical malpractice not to have an AI check a medical diagnosis and then eventually it will make it to check a governmental thing and then eventually AI will run the government how long that takes I don't know certain countries certain organizations may be faster than others but AI can already make better decisions than people across a whole Range of areas but like
I said the first thing will be the check and then the make what goes inside the those models then is important like we saw the anthropic paper on um model poisoning you know whereby with a few thousand words you can make a model out of a trillion words that these models have in them turn evil on demand and you can't tune it out and you can't identify it and you know these models are very sensitive like I remember an open AI Study people were noticing that open ai's chat GPT model was getting dumber and they
like are you downgrading this model and then someone did an analys that found out it was because the date string was passed into it every single prompt and in the winter people tended to be a bit dumber and do sadder prompts and outputs so that changed the whole distribution so when they changed the date on the computer it became smarter they are that sensitive So within that I Believe that every regulated industry Education Health Care government finance will need open source open data models and the question of who decides that data will be the most
important thing I think that we need to have open source to allow you to decide what data that is then transformed to knowledge and then wisdom by this contextual machine teaches your kids or informs your healthare and that needs to be a public good so this is where you've got the Decentralized open source kind of approach like how do we fund these public goods because the reality is we don't need a million education models we just need a Starship catcher level team to do it properly once and then update it every so often which is
also remarkable like you've got a amazing teacher in a box you have an amazing doctor in a box that's a finite level of computer and that's just part of this AI jigsaw Because like take for instance Healthcare your flops right now your compute that you use for your Healthcare is minimal compared to what it'll be in 5 10 years but then so are your hospitals and so are your countries and none of that can be exported outside that has to be on location and so again we need to build that infrastructure we need to decide
the governance of it and for me the best governance is permissionless innovation In that open source space that you can take and adapt as you want because the final part of this is that models are trained like graduates curriculum learning you teach a generalized knowledge then specialized knowledge then localized knowledge so we can actually just build that stack and that's kind of what we're doing intelligent internet wow that's one of the best Pathways I think I've heard you hear so much on this topic but it seems Like a lot of fluff like oh there's this
idea of decentralization but how do we actually do it that is the big question it's not being answered enough well I mean this comes this comes to kis theorem you know of organizational spread and metars Theorem of you know you can only have 150 friends like the friction costs of coordination are removed by this intelligence because let sa as a climate right why did Germany sh down its Nuclear reactors it was clearly stupid you know like that's Germany okay let's do call for the climate like it became very weird a lot of problems we have
are Global problems that require localized Solutions and coordination but we know have the ability to push intelligence down locally whereas now we can put a climate AI on everyone's phone that encourages them for climate positive actions whatever they may be you know or turns invisible people to visible this Is another reason I'm positive on Emerging Markets because our AIS will coordinate with each other better than we can but we better ensure that the ai's objective function is allowing us to have maximal agency versus trying to increase the profit margins of someone else because as you
tend to make Revenue you tend to become a bit evil particularly when it comes to this type of Infrastructure no don't become a bit evil as you make Revenue great Point well well it was that old do you remember the Facebook sadness experiment they took 600,000 users and they had the theory if we show the users sadder things they will post sadder things and so they made 6 300,000 people sadder and guess what if you see sadder things you post sadder things like you know it's like the YouTube algorithm let's optimize for Engagement which is
extremist videos and all sorts of things like that like we see it again and again organizations are slow Dumb AI you know like that are misaligned yeah I have I have a friend who says that capitalism is actually the first AI it's the first like Global coordination you know system system level intelligence and and that is misaligned because of all the perverse incentives and Market externalities uh that come come baked in with capitalism Obviously that's one person's opinion well I mean the organization was given rights as a legal person in many instances right and so
again they tend to move and they provision humans and there's a reason that people are sad it's mostly PowerPoints fault you know but other than that you do see misalignment at scale oh well before I hand it to Dave for some technical questions just one of the last points to touch on and I can't Encourage you guys enough if you're if you haven't read the article on X you should it covers so many great things one thing you spoke about was post scarcity economics so question for you do you believe that true post scarcity is
actually achievable through Ai and if so what do you see as far as the timeline on that well most of our problems are coordination problems we have enough food to feed everyone we have enough Ability to reach everyone and this is a time when things like solar power is getting down to less than a dollar per watt um we've got starlink we have other things I think it is possible to provide masses hierarchy of Need for every single human and a gp4 level AI to every single human basically for free um and I think that
allows us to coordinate into this human Colossus much better when you work at the math of it um but then it'll come down to like Why are why are we actually here you know these big question questions because there's a Japanese concept guy do what you like do what you're good at and do what you're adding value that value part is very interesting like Argentina had this uh program they Tri called the fs program where rather than giving Direct Cash handouts they had Universal basic jobs where the community decided on jobs and then that brought
lots of people into the workforce and Others and it gave them that meaning because I feel that if you just do the direct handouts to cover the bases like where is your progression and you'll see more and more of these movements uh you know it's Isaiah beran's conceptualization of negative Liberty versus positive Liberty like you're free to do what you want L say fair economics versus become part of this ISM you know more and more of those will go to the kids and say hey let's go And do something extreme and that's one of my
key concerns whereas what it should be about is increasing agency and community and we can use these AIS to do it like um one of my buddies built am my the AI you know based on the Reddit thing whereby you and your partner can dump a whole bunch of stuff and it tells you who's the from the argument oh my go got some ex users that would love that Dave we can solve so Many problems that way right and so I think it is possible but it just needs a lot of work but definitely we
have the resources to be able to do that and to light up the dark where we can't reach people well said well Dave I'd love to hand it over to you for um please you are guest that's a terrible there we go there we go great segue all right let's go um yeah so from a from a more technical perspective I just want to take first just a big step back to Provide a little bit more context and Framing and that is what so it comes down to capabilities and constraints and I don't know which
list is longer or shorter right now but what what is it that that you're seeing that AI can and cannot do uh as of right now and how do you expect that to change uh obviously our our our prediction Horizon is not particularly long CU as you said before we started recording exponentials are a pain in the neck um But but what are the constraints and capabilities that you're really seeing right now that that people are not aware enough of the capability is are just taking these massive leaps forward where narrow AI is just exceling
humans in every single area that I can see I mean this is like everyone's going to become Lisa doll the alphao player who like I can beat every AI in go and then boom he's beaten you know actually the RL methods real life for learning methods Are the same um and I saw that again you can see this especially with models like 01 and 01 Pro and R1 Now by Deep seek whereby the interaction modality that we had with this was like talking to a idiot savant like graduate student like Goldfish Memory there was nothing
there but it could just respond really quickly it could draw a picture it could make a song it could do all that now we have more continuous learning where it's learning about your preferences and with These new thinking models the key thing is what question to ask this a bit hitch highest guy to the Galaxy you know the answer is 42 what's the question right asking the right question is often you know it's prompt engineering and we've learned prompt engineering for human counts as well yeah I mean like my wife's been trying to prompt me
for 15 years now and she's succeeded we've see this all the time like so much is prompting in relationships in community The but this is a different type of prompt whereby again are you trying to create recipes and doing higher order thinking or are you just trying to get on with your life because we all think Siri's a bit rubbish but now we have all the technology for Siri to actually be a good personal assistant it just needs to be implemented and Apple's approaching that in the very cautious way but I think we'll get there
in the next year that's a capability thing where it's got The context and it has the memory because the way of view is this you have data you add context to data and you've got knowledge you've add a bit of experience and youve got a bit of wisdom there there and so we're building through these things and we're going from almost this filter which is a model because what happens with a model is that we take all this data and we can press it down and then it can guess the next word and that's all
it's doing or It can figure out how an image is destroyed down to its basic points in diffusion and then reconstructed so it's like a SE that you pushed stuff through but that's why the interactions were very quick there was all this continuous stuff to more this continuous learning predictable outcomes like again assistant type work so that's the capability side and again we're going into this higher order thinking now where it can take all these inputs and Figure out interconnections like with open ai1 model it doesn't show its reasoning if you go to deep seek.com
and you try that reasonable model and soon Gemini will be similar it actually takes you through every single step it's thinking like you say write some code to make a game of snake It's Like H what's a game of snake this is what a game of snake is this is how we should eat it and you can actually see it's working which is a remarkable thing in terms of Constraints it's we're not sure exactly how to use these correctly like again it's hard to build an 01 prompt it's still not the right types of interfaces
because we don't have anything that's continuous learning like for using an 01 type model again this Advanced model I should just be having a conversation with it and it should be learning and adapting and then writing the prompts and going away and thinking every so often we haven't figured out those Particular modalities yet but from a technical perspective to do the vast majority of human work much cheaper faster and better I think we're pretty much there on the robotic side it's just how many robots you're going to build like the hardware is there and the
software is getting better remarkably quickly and it's all being connected so the robots learn from each other you know from a visual perspective where it can see your Screen or outside you're better than humans already so all the component bits are there it's just putting them all together now that doesn't mean ASI that takes over the world we're talking about real economic impact at a fraction of the cost and this displacement comparative advantage question like I'm not sure what else we need to do to be honest given a certain level of cost which is the
cost of a human to do an Equivalent task right across just about everything I can see so that that actually tees up my next question perfectly which is uh you know deployment modalities integrating this with business and Society um and I know right now you're focusing on more decentralized deployment so what from a from a technical perspective what does that look like is it going to be like phone apps is is it going to be web apps um is smart home devices um what is it That you're seeing and and working on out there in
terms of how are we actually going to put like you know obviously robots are going to be one I think pretty ubiquitous modality um in terms of getting this out into society but what else are you seeing out there or what else are you working on uh so you know my Approach was open source so the at stability we built the best image video audio 3D contributed to protein folding all sorts of things right and we Had 300 million downloads of our models because we said anyone can download them you know and that created a
whole thing maybe kick this all off because then happen GPT and others so our approach is we're going to build data sets model systems and release everything fully open that anyone can just take and integrate I think that will cause a leap forward with aligned AI being core part of that but when I look at it I think there'll be four types of Intelligence there will be super expert intelligence that you call in every so often when you need to figure out something brand new but most people don't need that dayto day again the market
size for an 01 isn't that big you know it's like scientists it's theorists it's figuring out brand new things then there is Apple intelligence Google intelligence 10cent intelligence the intelligence that's just there you know and that can mostly be done on the edge And that can help you with your daily life and organize you it better it's so hard to find a good personal assistant you know everyone will have one everyone will have a Javis then there is llama style open weight which will be relatively Advanced and I think those will be the core of
intelligence reasoning systems so the complex systems are obviously open AIS and others and then there'll be these open source open weight things that are Infrastructure so I think the diffusion of each of those is a bit different and the time taken is a bit different but one of the remarkable things here is that usually an integration project those of who listening have tried that is just terrible right like you need so much handholding in human labor but now with ag which I identify as you know a level of autonomy and ability to get resources and
this moved to test time compute where it can take longer to Think versus this instant response you can put it towards a codebase and it can translate it from Fortran to Python and we're just getting all the specialist things just like a team together for that like we will see this with robots building houses better than anyone because again the communication overhead between these agents between these robots is so much lower so I think probably where we'll See it first is again anywhere that you see a remote worker you'll start to see those being challenged
especially at the junior level right because Juniors tend to be a bit annoying relative to seniors right like again don't I don't have time for that especially an increasing competitive advantage on the physical side like you know that's just a question of how many can you really produce only 70 million cars and 70 million motorcycles are built a year so I'd expect that ramp to be similar for that um but really like I said I think it's this outsourced area moving into insourced and then the modality will probably be a set it and forget it
just talk to these things to be honest like just treat them literally like an entity like one of the things I was thinking like if you wanted a real world human impact think about someone that you've trusted more than anyone in your life who's sadly Departed we can take their picture we can take their voice and you can talk to them on Zoom we have the technology to do that right now live and you can text with with them but that's our new workmates like literally we'll have zoom calls with them and we'll be talking
to them and right now they're not that great Devon and things like that but we're nearly at the point where they can be as good as The amount of money you put into it this again is the capital thing because something like Devon which is this autonomous software agent that's okay is $500 a month but you're trying to replace someone who's $10,000 a month if you put $10,000 of compute into a system yeah probably be better than your $10,000 person right or your $2,000 assistant or any of these other things that's a great point I'm
curious yeah so speaking of like okay relative Cost Advantage um obviously once these are good enough it sells itself it would be economically irrational to say you know what I can spend $10,000 a month on an expert you know developer or 500 or2 200 or $100 a month on a developer who works 247 365 so I'm curious what are the bottlenecks that we're that you're seeing out there in terms of um I guess there's kind of two halves that I'm thinking about one is development um you know is is the Data wall a problem is
it a constraint of computer or energy but then also bottlenecks to adoption you know we talked earlier you know user experience is is is still leaves a little bit to be desired but taking a step back what are the what are the bottlenecks you're seeing on on both sides from development and deployment I think this is about just putting together the Lego like again you got intentionally build it right and I think most of the Innovations are done this is the thing that's really terrifying like let's think of an optimized approach you go to emphasis
or BP company and they assign a bunch of people to you right and you never meet them how far are we from the point whereby that level of performance is all done by Ai and you don't know it okay yeah everything else is exactly the same you know like in the decentralized thing There's this whole thing around North Korean hackers like if you do enough hiring you'll encounter a North Korean hacker you don't know because it's decentralized and distributed the teams right but this is again how they get inside and they have all these hacks
but I think inin the next year maximum two you're at that point where by one of these business process Outsourcing companies you don't know that it's an AI On the other side and they're interacting with you completely normally you're even having Zoom calls with them and we again we have a technology not that's indistinguishable right now right and so this is the most natural way that I can think of the interaction which is just fitting within our existing systems and you think about the CV of that AI that can learn from all the other AIS
it's got a really great CV you know it's trained on all the Stanford Material It's trained on all the Harvard material like it's been to every University and it's learned from all its mistakes right I love it and you're like that's kind of we never seen anything like this before and it doesn't need any retrofitting this is the thing it doesn't need any additional capex on your side so we look at this thing the energy equation you like there's GPU constraints like anyone that's used the anthropic Claude models there's never Enough and you H the
capacity limits that is a constraint now but the scary thing is we don't know how much energy is required to do a unit of intelligent work well it's Al that's a moving Target as well because you look at you look at the the the exponential decrease of you know cost per token over the last you know 3 four years because remember they used to sell tranches of tokens in the thousands and now you sell it in the millions so does That am I am I correct in inferring that that is also a proxy for energy
cost as well no that's the scary thing a lot of people say that it is but I'll give you an example deep seek with their new model they have a 1.5b model that was trained off the quen model from Alibaba mhm distill down from their big reasona model their 01 equivalent in certain benchmarks it outperforms GPT 40 and it's 1.5 billion parameters to give you an example of what that means that would Run on a probably Windows 2000 PC on your smartphone right now that 1.5 billion parameter model will go at 100 tokens a second
faster than you can read and the cost of a million tokens on that is 0.0 one cents right something like that and the models are still improving so this is why I say we don't have a lower bound on where we're going to end up because the models are still full of junk they're still full of crap and they're getting like 96% reduction In cost overnight for a certain type of model we see these things all the time so this is the scary bit because how much can you do on your MacBook or your Mac
Mini or your smartphone does it make sense that your smartphone can code at the level of a novice programmer no utilizing the existing capex that's there no new chips no new nothing and again these models can work on machines that are 10 years old 15 Years old right so that's why we have to identify almost what is a unit of intelligence a particular task and then we can interpolate the cost of that particular task you know like um what was it um Arc did a chart of gp4 level intelligence it literally is a straight line
going towards zero and three years right yeah yeah so on on that topic so it it sounds like my inference was was Actually underselling it that that the the cost of energy is actually dropping off faster than I than I expected um I'm curious from the cuz you mentioned earlier and I wanted to revisit this topic was safety and alignment because you said that that uh that you know open source aligned models are actually a big part of what you're working on so I'm curious what do where do you see uh what what is the
state of the industry right now in terms of safety and Alignment Like what's top of mind but also what is the direction that we're going like what is what is what are the trend lines that we're seeing out there in terms of uh safety and Alignment in particularly open source models yeah so I was like probably the only CEO that signed that six-month pause letter because this is crazy and we need a break and the new aren't coming for six months anyway you know um people don't really understand alignment is what I found being amongst
All the big guys because aligning to what we can't even align ourselves right and you kind of again look at organizations you look at ethics you look at all of this the conceptualization of Ethics in China is very different to the conceptualization in the US very different to the UK very different between Judaism or Islam and kind of other things and we tend to say aligned to what we believe which if your corporation is aligned to your profits So if the objective function of the AI is not to maximize your agency but instead to get
you to use the AI and increase your token count and subscriptions it's probably going to end up misaligned in one way or another and that can lead to these weird things like I said YouTube optimized for engagement which optimized for extremism and very weird videos if you go down the YouTube rabbit hole you know you get like d the Explorer drilling Spider-Man's teeth and All sorts of weird stuff right this is whole subculture any of us that have kids will have seen that you know oh yeah it gets real weird out there my nephews and
and nieces are just like I'm like what are you kids watching and the weirdest thing is children don't use subscriptions they don't know how to curate their own information feed they just allow the algorithm to just feed them whatever I'm like what are you doing so yeah I've seen That so now you think about it you look at meta and Google their objective function is selling advertising but now they can have AIS that are more convincing than anyone you know Scarlet y Hanson's voice on steroids dripping with honey with a bomber on top and a
bit of Winston Jill constantly adapting to your mood faster than real time we can build that now and someone will build that you know and so that's Mass manipulation and so you're looking at That you're like someone should probably build open source models who the only objective function is to increase the agency of your child you know and we should have that as infrastructure because it has defined cost because we've got to build a data set that is common knowledge because any regulated industry could only use common knowledge because you can't have it and Bay
wall and then build those models and then deploy them so that anyone can take That because we don't want to have to decide what teaches your kid or what runs your financial services or runs your government eventually God forbid you know and we need to delegate that down to the local level to make those decisions and impower people to do so because otherwise you will use someone else's system to do that you will Outsource your cognitive intelligence as an individual organization a society um and that will be weird and Think a lot of this is
the data as well like you're talking about this data wall and oh we we need 100 million trillion tokens like the top models train on a trillion words five trillion words now most of the internet is garbage and most of that is junk you know like we've seen the Microsoft five models outperform because theyve made generative textbooks the textbooks are still a bit crap if you want to build an AI for an oncologist does it need to see more data Than an oncologist sees throughout the later part of their life no but no one's curated
those data sets and built specialist models because we were so focused on generalist polymath models that could do everything so there's so much junk in there it's like we using these big supercomputers to slow cook you know like souid crap steaks we got to have better food and brasset goes in there I like that analogy so that's actually Something that I was that I've been thinking about more and more because you know some people swear by 01 or Sonet or you know they everyone has their their own preferred model and in some cases it just
comes down to taste right it's like okay the the you use the tool that you know how to use best so do you see like a combination of closed Source open source specialty models general purpose models as you said earlier like the the Savant levels super intelligent models For certain problems are we are we heading towards like one model to rule them all or it sounds you're more like on the side of we actually need a huge diversity of models and data sets is that is that kind of a fair interpretation pretty much yeah I
mean again it's this individualist versus collectivist Vision right it's AGI God versus AGI is a complex hierarchical system and so for me like I think about how would I build an organization to do A task I would go with the people in front of me and my team then I get some consultants in and I get some experts in and most tasks devolve down to that but you got to know that these things behave in predictable ways which we now do through function calling you want to know the capabilities of these models sometimes it's just
a Vibe thing you know I really hate that person I don't get along with him whereas that one kind of gets me you know like it can be as Simple as that I think part of this is because of this Spectre of AGI that had this big thing which is again the first person to get AGI will use that AGI to stop all other agis it's called a pivotal action and then that means the Chinese can't get it you know you like again it created these weird game theoretic Dynamics versus thinking like what I really
care about is like I want my kid to be really well educated you know I want to know what that cough that I had is I want to have something that I can engage with to explore these topics a bit better I want to have a therapist who doesn't judge me you know these types of things that's a good one that's a great way to put it what do you think about AI generating its own data you know the idea of synthetic data is there any basis in that I mean it's better than most of
the crap you see on the internet right like we actually Trained a model on Reddit data and it got worse it was the worst language model that came out we were kind of surprised by that we broke the scaling law because we used it in sight the wrong way and like again rubbish in rubbish out to a degree these models do figure out inter relationships and interactions like I think the future will be these models iterating and improving data and then what's your minimum viable data when we were Justable diff Fusion it was like two
billion images went in that's a lot of images nvidia's latest SAA model use 25 million images for the same performance wow because how many images do you need in fact when we did an analysis is of the different sectors that lit up in the little neurons like 99% of the data was never used in 99.999% of the props wow you can even do crazy things like um you can delete layers from the AI model and it's still As good wow does feel a bit like Homeopathy sometimes as you distill it down where does it all
fit like because like a seven billion parameter model can write all of Wikipedia wow in better than Wikipedia right like again you can see it you can just try it like ask it about any Topic in Wikipedia and it can give you the knowledge you know Concepts but then Wikipedia compressed is like 26 gigabytes or something like that you're Like where does it fit right you know that that reminds me it's you know distillation quantization and in this case it almost sounds like you know neural pruning or lobotomizing the model um it it is almost
do how do you how do you mentally think of that is cuz the way that you just described it it's almost like the knowledge becomes holographic not even compressed it's like a hologram is actually what is stored in the model weight so do you Think that's an appropriate way of thinking about it or is it just another kind of of embedding compression what we learn through school is principles right like okay you might know when Napoleon died or something like that but mostly it's these principle based things where we're compressing intelligence down to principles in
order to react and adapt and so if you look at what models do the latent space it is a multi-dimensional kind of individual Thing whereby if you have a image generation you're like cup and it's cup your ears cup your hands well cup it all depends on what's around that so I think the actual knowledge compression level is Tiny and it is kind of holographic but we don't really understand how small we can get and this is the thing we're still in the research phase where we're just chucking stuff and applying a big hammer we're
not optimizing at all which is crazy and that's why I said these Extrapolations of all the energy in the world and a Dyson Sphere will be used for this and we'll surround the sun I don't buy that because I don't know where the lower bound of energy for a given task is well I know that it's collapsing and I have no mental model for how far that collapses what is the capability of a model that can run on an M4 MacBook if you hyper optimize it nobody knows because we haven't even got there yet we
haven't got to the proper Engineering thing where every single tiny bite that goes in every single part of the process is optimized that's the next few years and I think that capability will be better than 03 but 03 is better than all but the top 10 coders on code forces right and it's better than just about any lawyer and like but that fits in a consumer Hardware you know which is insane my last question uh is is more just on the personal level because now that we've Had a I've gotten to a Tipping Point personally
where I've been using AI pretty much every day for a year straight um you know before that it was tinkering you know gpt3 and fine tuning and stuff but now that if I have several AI empowered apps on my phone and you know on my computer I find that the AI is training me as much as it has already been trained and it's there's actually many cases where I need to use the AI less and less and because what you said Earlier sometimes it's a skill issue right sometimes you just don't know how to use
the tool sometimes you don't know how to prompt it but eventually you learn how to use the tool so there's there's an acclamation period for for us as individuals um and so my question then is have you have you w experienced that same thing personally um or or or have you seen it out there in the world like maybe a a big chunk of this is just it just takes time to learn to ride this New bicycle pretty much it's like a rocket ship for the mind right like it will upgrade you we saw this
again I think probably the best study of this is go again this Lisa doll Alpha go thing the average level of professional go players in the world went up and you can see that VI the ELO scores because they figured out new ways to play in their brains and a human brains still the best figured out new ways to understand the capabilities and Constraints again this is what you do when you have a team this is my individual capability I know your strengths your weaknesses I can do more now and then we go out and
we split the atom or we freaking hopefully go to Mars or whatever and so in doing this you know what you can do this is why again like I recommend maybe make it hard hiring press I don't know everyone should spend 30 minutes to an hour just trying Cur Right and getting to know and the first time they build their app like I can build an app like can you build the best app no but if you keep going you will be able to you could didn't think you could do that before and your brain
will be S wiring a different way where you're like well that spoiler plate I don't need to do that anymore right similarly the other one I tell people to use is notebook LM like the ability to generate a Podcast based on up to a gigabyte of stuff on your Google Drive and then you can call in and talk to the hosts to discuss it it's just such a natural way of doing things and people don't realize that's the capability but then they're like when I'm trying to understand a piece of knowledge this will be my
default way of doing it but maybe I don't need to do it anymore because I'll approach knowledge in a different way because my tool and Capabil is increased this is what I call the agency versus a gentic divide again a lot of the western Labs T Silicon Valley labs are like how can we replace humans versus let's upgrade humans again just like we did with the spreadsheet with computers with mobiles again this cybernetic kind of thing where information is portrayed to us in the right way and then we can make decisions based on that mapping
um and so I've s felt that myself you know like again I Can do a lot like it's a lot easier for me to write a book now or write an article or things like that I'm just shy to release them you know like that you know like my art was bad now it's okay but I'm still too shy to release it that's a great Point good question Dave to add to that you know I wrote I I'm very transparent that I use Ai and everything so I used AI to write a keynote and I
told everyone when I was on up on stage I'm like I didn't write Any of this you should know that AI wrote Every Word so there was somebody that came up to me after and he's like I'm not offended because she told us it was AI but there was some content in there that came straight from one of my books and I was like oh my gosh and then I realized that's the nature of the LM Beast where it's like you know pulling that data from someone else's thoughts and the way it presented that data
from his book was in a very original way so He ended up it felt like a completely new conversation to have because the world I come from is like copyrighted writing materials that are highly they're behind pay walls I've had a writing agency so the conversation was extremely positive to where he was like I really like how AI presented my work and because you told us AI wrote your keynote I'm not offended well I mean again this is this is going to bring into question so many Things like IP wrs like why do you have
patents because it's hard to do coordination and exchange ideas seamlessly right whereas I can dump stuff into my eye they can talk to youri and then we can jam then we can come up with an invention right like that's crazy versus again the pace of these things the cost textualization thing he liked it because it was in context and a lot of this legal stuff is going to face so many Troubles because like when you have texal Optimus robots walking around are they going to cover their eyes and their ears if they see a can
of Coke you know or listen to Taylor Swift stop training the new Nets it's probably not going to happen right like what happens when we have neural links you know even Rayband glasses like you know the next generation of meta glasses has this is a crazy time when our capabilities get so much ultimately it's About information in the right way presented to the right people to make the right difference y actually something I've realized recently Dave this is probably interesting so I have our Fantasia I can't visualize anything okay yeah know so see yourself on
the beach do that in fact probably 20% of the image development devs have that versus less than 1% of the population once I started using the really fast real time stable diffusion stuff I Started to see color for the first time so my brain is actually rewiring itself and that was the craziest thing so probably need to fund some studies in that but would not surprise me at all w we had this we had this paper called Mind ey did you guys ever see that where I don't remember I don't think so tell us about
it real quick so you take functional MRIs of people when they're seeing a can of Coke or something and then you reconstruct the can of Coke From the functional MRI yeah I have seen that okay yes yeah so the internal representations of the mind I think these feedback loops and just that pace and the ability to adapt you got to be rewiring something yep I remember the very first time I sat down in front of stable diffusion in mid journey I was like I had such a hard time explaining what I wanted and I actually
realized it was because I had lost the ability to build scenes in my in my own head which I did have that when I was little cuz I would sit there and just fantasize and I would like design things in my head I was a very visual thinker but then as an adult I became more of a like a a like code based thinker or or word or symbolic based thinker so I've actually kind of lost that ability I don't have like full a Fantasia but I know what you're talking about and I've gotten that
ability back because now I'm thinking more like a cinematographer Like oh yeah here's exactly how I want that thumbnail to look and then so not only imagining like what the final product will look like then figuring out how to express that to the machine so I think you're right um that that we are retraining ourselves with these tools it's just like video games again you look at the response patterns of this your your sub 50 millisecond one your 500 millisecond one your planning one you think about the Arc of Mastery in Video games it's actually
the same thing and so again that's how we adapt as humans and then it's just a question of right now probably what a thousand people in the world are using AI to its fullest extent that would be my guess a thousand out of seven billion right so if you just dive in and you just think all the ways you can use it and you try everything you Tiner like you'll be in the top 1% .1% .0 one% yeah that's the first mover's Advantage that we're working there go I want to I want to just reflect
on and amplify that last thing that you said because number one we're nowhere near the thermodynamic limits of computation let alone the optimization and and distillation that that's capable in these machines just from a first principal's energy perspective but then also we're so we're still super early in the game the level of saturation uh is still very low and that always just Makes me feel so much better and that's why Julia and I got together on this Mission that's that's why we picked the name first movers because that's who we are and that's who we
serve and our mission is get this stuff deployed safely and effectively into society as much as possible all right that's my Spiel there you good it's exciting times interesting times please keep sharing all your thoughts on X too you're doing amazing work Imad to Drive the future of not just you know you built something incredible and you stepped away to have a bigger conversation and I think where that goes with how the world response to what's coming is literally the most important thing we could ever talk about so keep talking about people talk about that
thank you let's get more people talking about it you know like that's why we're here that's the whole point Y yep so thank you thank you So much it's been wonderful to meet you and and talk for the first time and have this conversation so just personally thank you so much it's been great yes pleasure thank you you're just as smart as you seem it's like you don't know until you interview someone it's the it's the AI in the ear just whispering to me no have you seen actually there is that thing um that a
lot of people using it does actually like record stuff and now Gives you live tips on how to be smart as you're do many people need that yeah this this is thing it's gonna be a crazy time going forward crazy time absolutely such a privilege such a privilege hey thanks so much for listening that is a wrap on episode 10 of the leaders of AI podcast it was an absolute honor to have imod mustto the founder uh and CEO of intelligent internet and the found and former CEO of stability AI he is by far And
above one of the people that absolutely spearheaded what we have in artificial intelligence today especially with generative AI his company opened the door to multimodal Jin Ai and it was amazing to be able to sit down with him and see what he sees regarding the future definitely go check out his Twitter handle which is just simply EMA stock and you can catch that article we talked about as well as all of his thoughts regarding the upcoming labor And capital divide if you haven't subscribe to our podcast be sure to do so leaders of AI podcast
with myself Julie McCoy and Dave Shapiro is streaming on every podcast platform and of course YouTube where you can catch the video versions we look forward to having you back here on the next episode see you then
