Investing in AI, Crypto, & Tech in 2025 | Elad Gil

I do think that to some extent one could argue that there will have been three ages of humanity right the first age is um all the compute is roughly humans and then other animals the second age of humanity is probably this what we have right now which is there's some split between humans and machines and humans are directing the activity and all the rest and probably the third age is the age of machine intelligence right where that's the predominant form of compute and Intelligence on the planet

welcome to bankless where today we explore the frontier of tech investing this is Ryan Sean Adams I'm here with David Hoffman and we're here to help you become more bankless I led with tech investing I mean crypto of course is a form of tech investing we're talking a little broader than that we're talking not just crypto we're talking about AI today we have unicorn Tech investor elod gil on the podcast he's giving us the Silicon Valley take on AI crypto and also silicon Valley's involvement in politics particularly the recent

involvement so this is one part getting up to speed on everything that's going on in Ai and uh elad provides a master class in that and then we talk about the intersection of AI and crypto and his take on that and then politics how will Tech flourish under the new Administration what changes have we seen in the last year Eli probably knows as much about AI as Ryan and I bankless knows about crypto uh and vice versa he knows about uh crypto the about the same amount that I think we know about AI so

we really start off picking his AI brain uh but he's uh just familiar with the growth of markets the growth of tech having seen the rise of the internet uh and so not only is he intimately familiar with these Tech sectors especially AI but also how investing works and how markets develop and grow and the s-curves and how all the these many s-curves uh relate so feels a little bit like ancient wisdom from a tech investor veteran on the podcast he's also very calm which I found found very peaceful uh so overall very

enjoyable episode does he meditate David usually call it when guests meditate and you can he could he could be a meditator he might meditate he might be a meditator we didn't get into that in the bulk of the podcast I my big question going in for him was like okay give me the truth on AI is it overhyped right now or is it appropriately hyped or is it underhyped so stay tuned for that answer as well yeah let's go ahead and get right into the episode with elad Gil but first I want to talk about some of these fantastic sponsors that make this show possible bankless nation very excited to introduce you to elad Gill he is one of silicon Valley's greatest of

all-time investors uh I would say that he's he's backed 40 unicorns including Airbnb our beloved coinbase figma stripe many others he's super active right now in AI he actually hosts a podcast that I enjoy very much it's almost like the uh the AI sister podcast to bank list it's called No priors uh he's just in general a very preent thinker today we want to get his takes on AI on crypto on the next decade of tech investing what the opportunities are elad welcome to bankless oh thanks so much for including

me all right so um let's start with AI because David and I have been following this in various ways the bankless audience uh knows about it but you like we we focus mainly on on the crypto Tech Revolution not as much AI um catch people up catch us up to speed so obviously there's a lot going on right now how about this year if we just zoom into this year what are the big things like the big milestone events that happen in AI that we should be paying attention to as general Tech investors yeah you know it's kind of interesting because if you look at the history of machine learning and Ai and you know

weed to all call it machine learning you know 10 years ago I think what a lot of people really underappreciate is that we had a series of fundamental breakthroughs that effectively put us on a different technology curve from what we traditionally talked about machine learning so we used to talk about convolutional learn networks and recurrent neural networks and all these things in the 2010s um with the Alex net and a few other breakthroughs around Machine Vision and um you know a few other areas and the basis for those things kind of shifted in 2017 when Google invest invented what are known as

uh the the Transformer it's a a specific type of architecture model for machine learning that got implemented at Google but also to openi and the T and GPT is Transformer and it wasn't really until chat GPT launch just two years ago that a lot of people woke up to this fundamentally different technology curve that we're on so it's almost like we're on one curve with machine learning and then all this Transformer based stuff came out and it kind of boosted us into a different trajectory and honestly a different technology curve and the capability set is very different from what we've experienced in the past with

traditional machine learning which is really effectively running a bunch of regressions in some sense or kind of data mining out statistical correlations between things this wave of AI which a lot of people are calling generative AI is um about the ability to understand and manipulate different types of language and imagery and a few other things in language includ things like codee and language includes you know the synthesis and understanding of knowledge and there's these multi-step processes that you have and all these other things and so when people talk about AI we've been talking about it for literal

decades but I think the specific flavor of AI that we're focused on right now is only 2 three years old so even saying what's happened in the last year is almost like saying what's happened in a third or half of the time frame of which we've been aware of this really interesting technology shift that we're undergoing right I think there's probably a a mainstream interpretation of Ai and how it's going to impact Our Lives we see chat gbt we see its implications um then there's stuff like you know uh custom AI generated videos and you know deep fakes and this is all

going to impact our lives in this particular way there's probably a consensus level understanding of how AI is going to impact us but I'd actually like to uh see if there's any difference there between what you think is the future trajectory of AI and if there's a gap between mainstream understanding of the future AI and what you think is true is there a Gap is there an alternative uh version of ai's future that you think is being under represented or under discusted that uh mainstream Society mainstream Tech forward people are maybe missing I don't know if it's under

discussed I do think in general AI is still underhyped despite being incredibly hyped right and I think the the thing that people fundamentally misunderstand is as we have these advancements over time really the product that you're selling or the end product of these AI systems is units of cognition right you're selling pieces of thought or ability to do things so for example you look at a company like decagon and what it's doing is it's augmenting customer success agents so you're a customer support rep and you suddenly are making that person's job dramatically easier or you're tackling

more of the queries that they get through users and so each person can handle a much larger base of people and suddenly a person who only speaks English can support people in 30 plus languages 247 you know you're kind of Shifting the Paradigm of how work is done and The Leverage that you're getting on work and so for the digital world you're B basically selling units of labor for the physical world if some of these advancements in robotics come through and you know I think that's uh we're much earlier in that curve then eventually you're selling units of robot

time or labor time or however you want to phrase it and you know it does seem likely that one of the major types of robotic chassis or approaches that we'll have or form factors is going to be humanoid um a because the world are designed around that and B because it's general purpose purpose in terms of the things it could do right but that's that's much further ahead in the future if you just look at the digital part of it which is very clear right now I think eventually you're selling units of Labor or thought and that's very different from hey we're doing this regression on a bunch of data you know which was machine learning it feels uh a little

cloudlike as in there's a cloud of Labor out there that you can purchase units of Labor is that an acceptable comparison yeah I think that's a great way to put it so as an example eventually you'll have a series of bots that will do aspects of coding for you and already people are using coding tools like cursor or magic or or um Devon or other things right but fundamentally as those capabilities get better and better you're going to have the um AI system to write more and more code for you or the I system do more and more customer support for you I don't know if you saw

the um tweet from the CE of Clara that came out maybe six to nine months ago where he said that they let go of 700 people on their customer support team because they replaced it with something they built on top of w open Ai and it had a higher net promoter score they had a 25% reduction I think in repeat queries it was available 247 and I think it was like 20 plus languages and uh the time to actually resolve the issue for a um a customer went down dramatically I Remember by how

much 50% or some you know significant amount so strict Improvement across all domains the strict Improvement across all domains with an AI system right effectively to your point on the cloud they they created created a cloud of customer success agents or help right now it wasn't quite agentic agentic almost implies like this thing is going to be self-acting in really deep ways and the technology basis still has to develop to get there but again I think it's back to it's units of cognition or or labor or however you want to phrase it and it's units eventually of robot minutes or robot time or physical labor time or and that's eventually what this

wave is about right at least for a language right there's image gen and video and then there's all these Foundation models being built for phys and Material Science and biology and all these other areas which is different and um we'll have different substantiations and kind of overlap in terms of um what type of person they're they're going to augment to rep place what I think uh I I would love to find out from your perspective elad is like where we are in kind of the the scurve of this new of this new Transformer uh sort of unlock that we've uh approached in AI so you're

talking about the measure being units of productive intelligence and and the case that you just gave is basically like uh some Corporation has a you know customer support Intelligence Center that is primarily staffed by human agents right now right so those are its units of of customer support productive intelligence and what it's just done is in using you know open AI it's it's been able to replace that productive intelligence with a higher form of productive intelligence right so you see that that could happen in Customer Support maybe

some white collar uh type jobs we're also getting these version upgrades with uh you know the chat gbts of the world right so like we go from three to to four and to like five to six but at some point in time I mean we know in crypto right you sort of you're on the other side of that scurve and kind of adoption progress for the the technology that your curve that that that you're on it starts to PE Peter out I mean where are we on the curve here is this still early or like do we have a lot of upgrades to go or like when does this uh version of

the tech curve kind of uh stop and uh start getting diminishing returns yeah this is um an area of active debate within the AI community and some people say well uh and and really if you look at AI there's three or four components that go into how smart the system or how capable the system is and part of it is how much data and what sort of data do you have to train it and how clean is it and how is it labeled and how is it generated and etc etc um there's how much computer you throwing at the thing and you know

uh there's a lot of fine-tuning or posttraining what do you do once you actually have the system up and running and then there's the how much computer are you actually allocating at the time that you're doing inference which is sort of the moment that you ping the AI system and ask it to do something on your behalf and each one of those things can scale in different ways right and so um 01 from opene is really focused on that last piece and you see that a lot happen with people like you get asked a question and if it's a hard question you'll think for a minute or a few seconds or whatever and then you'll answer you won't just spontaneously answer and an easy question you

immediately belt out right and so differentially allocating compute is what you do as a person right you're kind of thinking more or less and similarly that has its own scaling law that people think has a lot of runway on it the training side probably still has quite a bit of Headroom on it in terms of just the core I'm going to train a big model on a a bunch of gpus and on the data side there's ongoing questions of like do you move to synthetic data are you capturing data in new forms are you going to specific types of experts for the the post

training side of it so I think in each of those there's debates around how much room there is but I think overall there's an enormous amount of room still improved so on the one hand I think we're still reasonably early in the scurve um and that's again a debatable topic but I think there's a lot of Headway just with the stuff we're doing there's a separate question of say that we stopped all progress and we just took the models and capabilities we have today how many more applications could we cover for it and there's a ton right again this technology in some sense or people's awareness of the tech is 2 years old GPT 4 came out I don't know

what 16 months ago right and four was a big stepping capabilities versus three and that's when you could suddenly do legal so you know one of the companies I backed is called Harvey and they have this sort of Legal Assistant in set tools and um on 3.5 which was a was the version of gbt right before right before four they couldn't do legal workflows it just didn't work it wasn't smart enough and it forced something it was capable enough right and so um I kind of call it the gbt ladder step right as you move up from four to 5 to 6 to 7 you're opening

up entirely new markets that couldn't be served before because the thing wasn't smart enough to serve them or you're fundamentally changing the capabilities that you can do in those markets and if you look at um SAS in the US it's a half trillion dollar a year stas and enterprise software it's a half trillion dollar a year Market if you look at all the sort of white cish Services the payroll for those in terms of areas that could be impacted by AI is about 3 and a half to 5 trillion so if you convert 10% of that headcount cost just employees salaries into SAS

revenue for AI you've recreated all of SASS and Enterprise software in terms of market cap right so this is a huge Revolution and that's why I'm saying it's underhyped right and so there's there's two s-curves it's a technology s-curve and it's really a few different S curves stacked on top of each other as mentioned there's the training and the post-training and there's the time of inference stuff and then there's the s-curve of adoption and on the S curve of adoption really early we're basically in the era of you know it's a year after um the Bitcoin white paper dropped or

something and we're about to have like the mount gox blow up or you know whatever analogy you you know it's really early do you know what I mean yeah okay so you're painting this portrait of a stacking SC curves all of which are early including the adoption uh scurve here c c can you give some insight into how transformational this will be so you've been in Silicon Valley since the early days you've seen kind of the the birth of uh the the internets and you've seen mobile and you've seen you like crypto and and you've seen social media you've seen all of the these kind of Trends right and famously

Peter teal talks about how like um some of these uh you know software value you know creation hasn't really filtered down into like real world productivity for the rest of the anyway Is this different or how does this compare to previous revolutions I've heard some people compare the AI Revolution this this idea of unlocking productive intelligence to the Industrial Revolution which was kind of like a whole new scale of productivity output for Humanity and I think that's a different trajectory

maybe than what the internet was anyway how does this Stack Up in terms of the the tech that you've seen come out of Silicon Valley since you've been there yeah I would separate out two things because again there's the digital Revolution and then there's the atomic revolution of Robotics right which is again we're not quite there yet on the second one um but that has its own transformation curve or it's its own applications sort of globally and um self-driving cars is sort of one substantiation of that in some sense too and that's coming the atoms in the bits and you sort of separate that I would separate it out yeah because it has very

different Society level implications but also um the degree to which you re rework cities and physical labor and everything else is fundamentally different from you know what what can you do with digital information and the bits is happening now that the bits is happening now yeah the bits is happening now I mean the internet was the biggest revolution of them all on for this stuff because a this AI these set of AI systems couldn't exist without the internet and distributed compute and neither could crypto right and so I think a lot of the um a lot of the

really big technology waves were just outgross to the internet in one form or another and so that to me is sort of the the granddaddy thing or whatever you want to call it um but on a relative basis I think this is much bigger than mobile um it's going to be bigger in some ways in certain aspects of social uh crypto I think has been enormously transformative uh mainly at least today in you know the financial sector and obviously spilled it over into art with nfts and you know it's spilled over in other ways but I think a lot of the kind

of web three premise of everything is going to be on the blockchain you know have airb be on the blockchain with a token to incentivize host you know that stuff hasn't really proven out quite yet um and so I think uh you know it's going to be a very big Revolution and it's going to take a decade plus to substantiate one thing that we've seen in crypto is kind of these repeating hype Cycles where every sort of four years or so crypto goes out of tear and then it sort of gets ahead of itself in terms of you know uh the market price reflecting kind of the reality of where

the technology and where the adoption actually is sure and I'm trying to like paint this picture for crypto investors onto AI do you anticipate something similar where basically it will happen in waves and you know I don't know where we are in the first wave but it's been incredible to see the evaluations of of companies like Nvidia of this cycle just like you know absolutely explode into the most valuable company in the world and so it's hard to look at that and not look at some crypto analoges well okay maybe AI is going to be a long-term

transformational technology but are are there periods of time within the Decades of that transformation where it gets kind of overhyped from a market perspective like are these companies making money do they have business models yet yeah I mean Nvidia is clearly making money so I think that's valued so highly but that's one of those like reflexive money type things right it's like it's it's making money because there's so much demand for uh gpus and chips because there's so much you know I guess hype going into the the other elements of AI well it's not just hype

it's really interesting so so if you look at the foundation model world or at least llms these large language models which is you know um open AI GPT or um you know Claude from anthropic or sonic from anthropic or you know kind of name your your model uh Google and B and all the things you're doing there um or or Gemini um you know fundamentally um the reason the hyperscalers ended up as a primary backers of these big model companies right AWS probably is the biggest backer up in Tropic now and

Microsoft is a biggest backer of open AI Etc is because it also drives enormous revenue on their clouds for AI Services right so Microsoft had something like a $28 billion quarter last quarter and I think they publicly said that 15% of the lift on that quarter that's what three and a half four billion doar came from AI That's that's incredible that's that's significant right and so these things are translating into real revenue and it's happening at startups where suddenly you see a startup go from 0 to 10 to 50 in two years which is insane in

terms of any sort of traditional SAS application you see that in terms of the rumored numbers around open AI where they're now in the billions of Revenue after two years for three years of offering gptt as an API that is actually being used right I mean these are insane adoption curves now there's a separate question of what is the durability of a given company relative to these adoption curves because overall the segment's going to happen it's such a useful and Powerful technology and it gives you so many capabilities and so many cost savings and so many new revenue streams

and all the rest of it that it's happening and it's going to happen and then the question is who wins and for each layer of the stack right you have the foundation models actually have the chips with Nvidia you have the foundation models you have apps actually have infrastructure then you have apps and within apps you have B2B and you have consumer and within that stack of stuff for each sector you can kind of go through and ask is it an oligopoly Market is it Monopoly is it highly fragmented who wins why what's the defensibility of each one of these things right what's the technology basis for winning do they have to build their own models or not so there's all that

stuff right and so I think the direction is clear the who in some cases is clear and in some cases it's less clear and it's the old thing about how the future in some cases is determinant but you just don't know who's going to be the person who drives that piece of the future right but you know that future is coming and I think that's very true here I mean we've definitely seen that with with crypto it's been hard to predict who the individ what the individual networks that like win out are going to be but we sort of know that the future is inevitable um I I want to ask you a question about the evolution of this

market so you're painting the picture elad of this is an early Market it's still actually underhyped from your perspective even at this point in time uh how how do you think this industry evolves and there's different lovers on this like a lot of the the value right now seems to be going kind of I don't know if You' call this the platform layer but like you know uh the mag 7 and some of these big companies right there's also this other uh layer we could talk about like closed Source versus open source I mean something that the you know crypto Advocates are very passionate about is um decentralization

right anybody being you like able to use this technology and there are different versions of that in uh AI but that's something that uh I'm sure you support permissionless decentralized the ability for body to spin up these tools and for not to be cloistered in some Walled Garden anyway what do you think of this this Market structure will it be centralized will there like be kind of power law winners here or will this be more more diffus I mean will we see Deca unicorns in the startup World start to uh compete against some of these mag s

companies yeah there's a ton of questions in what you just said so let me try and uh tackle them one by one I think one is around uh open source versus closed source and obviously I'm a huge fan of Open Source software um I think both will happen in this market segment and I think arguably both have happened in crypto right you have dexes but you also have centralized exchanges and a lot of the things that are supposedly decentralized are actually way more centralized than people really see it first blush right like how many people can actually commit uh to bitcoin core and you know how many

um miners actually make up What proportion of the network you know it's pretty centralized actually in some ways so uh or you look at on our other you know protocols and you know relatedly sometimes there's more centralization or less centralization right um and so uh the same is is going to be true in this world where the really big open source models of the foundation model layer at this point are basically llama from meta and then mstr right and there's a bunch of other stuff but at least for um the

language model side those are the ones that are I think most prominent and then there's other types of Open Source model for a a wide range of other areas in terms of you know things that have come out of Academia for robotics or uh weather simulation or biology or you know so you can kind of go through one by one and ask will it be closed Source or open source and to some extent um the hard part for open source and traditional software which is different from crypto where you can often monetize in other ways through a token or you

know there's two or three ways actually that you can you can monetize open source in crypto many of those things don't apply in the AI world and it's much more like traditional SAS where you have to um figure out a business model around the thing and charge for it and so for llama um there's a few things that I thought were really clever that Facebook did one is um if you're over a certain user number I can't remember what it was in the original license was like if you had over 700 million users you had to pay for it or license it otherwise you could use it freely and so

that meant if you're a hyperscaler and you were trying to put llama on your platform you had to license it or if you were one of the really really big social networks or companies with enormous numbers of users you'd have to license it everybody else could use it for free I thought that was very clever of them um can you elaborate on why that's clever why why does that work so well why is that a good mechanism um because I think it does two things one is it uh potentially and I don't have any insights into you know how Facebook thought about it um it potentially creates a monetization path for llama because all the big hypers scalers if

they want to adopt it they have to license it and pay for it and again if it's driving cloud services they should benefit from that it also I think means that the very large competitors of meta can't just adopt it and compete with meta using their own technology and so because then they kind of bridge the gap of having something that was roughly fully open source but for the things that I'm assuming they really cared about it was effectively close force or at least you had to license it right I thought that was really smart um and that me any developer can just pick it up and start using it almost anywhere in the world that's amazing right so uh I think

uh a lot of the value of that type of Open Source is probably going to go to the infrastructure providers and then app companies that want to use that as a differentiator the reality is at least today A lot of people are using um open AI or claw as sort of the starting point and then if they figure something out then they'll go and maybe fine tune and open weights an open source model like llama but it's kind of like tried down an API first where you don't have to do a lot of heavy lifting and see if it works and then if you think you can really optim it or you're worried about

data security or whatever it may be and again I think the data is quite secure using these other apis but if you if you have some concern about sending data back then sometimes you go down the open source route you know and find t a model or do whatever it just you need to do we've discussed the multiple uh adoption curves S curves of the growth of of AI and Y it seems to uh be that it can accelerate very quickly in the near term uh I want to introduce one more adoption curve as curve of technology which is crypto and we're I think we're starting to approach crypto like middle of the

scurve these days especially with Bitcoin Crossing $100,000 um I'm wondering if you're paying attention to the intersection of crypto Ai and this intersection has been growing uh right after chat GPT launched I remember there was like kind of a a surgence of crypto AI tokens like right afterwards kind of just writing The Narrative of AI and it was it didn't really make any sense it was kind of like I'm going to call it Stone Age version of crypto AI is very rudiment wasn't it wasn't real uh but that was

over two years ago and since then I think there have been some developers who are really trying to make this work like trying to figure out how do these two parallel Frontier Technologies how do they intersect how do they grow together and lately there's been some uh very strong Sparks that have probably actually turned into at least a small to mediumsized Wildfire in the crypto world hasn't really broken out into mainstream uh and this is kind of the story of uh

truth terminals I'm not sure if you're too familiar there's uh but the number one uh GitHub down uh forked and starred GitHub repo right now is the Eliza framework which is allowing people to build their own AI agents some with crypto wallets some just vanilla agents so I'm wondering are you observing this space is this space interesting to you and if you just have any takes yeah I have like four or five comments on it I guess um the first thing that I think is fascinating do you know that of near the near protocol you're talking about with IIA and team

uh I mean it was an AI origin right um yeah exactly IIA has a very well-sited AI paper yeah he's on the Transformer paper he's on the original Transformer paper he's the last author on that paper oh wow fun fact I didn't know that actually yeah yeah and so when my understanding is when he when he left a start um near and it was called near. a right it was an early AI company and they were originally going to do almost like um GPT style stuff and they decided they needed to do and he'll probably correct me in all this um my

understanding is they decided that they needed to do a lot of data labeling and they're like how can we pay people to label data around the world maybe use ethereum and of course ethereum wasn't scalable back then right that's why we have all the L2 stuff and all the all the roll all the stuff we're doing on top of ethereum now so they they said Hey how do we create a really scalable protocols so that we can create tokens that can be used to pay people around the world to label data so that we can then build a giant AI system I think that's the origin of near which is fascinating right mhm um or at least

some version of the origin story that I've heard of it um so I think there's long been an intersection of the people who are interested in Ai and the people who are interested in crypto and I almost feel like in a given era when one of them has been hotter than the other it's tip people's career in One Direction or the other you know I don't know if you know Umar Roy from succinct right on the podcast yeah she's great right and so that's a great example of somebody who did lot of AI and ml at MIT and maybe if she' started her company two years later she would be doing AI stuff right now I mean she's brilliant right and she's very smart on ZK and the

mathematics underlying it but you know there's stuff like that that I think is just fascinating in terms of these paths that are almost like Moment In Time dependent on when you graduate and maybe she would on crypto anyhow I don't know I'm just saying like she had that background right and so I think there's a lot of overlap in the backgrounds of people are interested in these things at least a subset of people so I think that's one aside just in terms of the the human capital or the people who who are excited about it I think um there's three or four approaches that people have been taking to that intersection of AI and crypto and a lot of it is over time has been can we create these data

repositories where you get paid in tokens to label data or use data or do data Etc there's the distributed compute stuff which I'm more skeptical about there's reasons you centralize these things usually one could argue worldcoin isn't part of proxy on Sam Alman you know mically um there's um identity which I think is super interesting and actually do you think like a block resident identity could be used by the agentic AI world and I'm still surprised nobody's built like a truly good identity system on the blockchain there's payments which is kind of the obvious one which is why maybe you mentioned you know building AI with the

wallet attached because it'd be natural for an algorithmic agent to transact using crypto um I think the censorship and censorship resistance is super important aspect of crypto that AI doesn't have and I think that is showing up in almost like what you call the politics of the models right because the models are beinging steer down very political the specific subset of political paths right they kind of reflect the Bay Area and the Bay Area politics is basically what you see when you interact with one of these systems and that may not be a good thing