Eric Wall - Crypto Tokens in the Age of AGI - TOKEN2049 Singapore 2024

uh I want to do a show of hands as well so in this room uh who here thinks that we are in the beginning of an intelligence explosion okay that's I think that's slightly more than half the room uh let me ask the other way who here doesn't think that we're in the beginning of an intelligence explosion okay almost I I saw one person raised their hand so only one person is is willing to say NOP NP intelligence explosion not happening um so if you raise your hand and and and you think that uh we're in an intelligence explosion and you're into crypto then this talk is for you because maybe uh you haven't fully real fully internalized what it actually means uh for yourself for your job for your coins for your kids for your spouse for your family every person you love every person you know and your very existence on this planet because even if you get that we're in an intelligence explosion playing that version of reality out and adjusting for it is not an easy task and there's a big difference between intellectually understanding something and emotionally being prepared to act on that belief um it's kind of like a chimpanzee that sees a human realizes that that human is is is very smart but it's very hard for the chimpanze to sort of imagine that that human is now going to fly spaceships go to the Moon uh create nuclear weapons and eliminate forests and change the ecosystem so regardless if the CH understands that the human is very smart and he's here now how is this chimp supposed to plan for that so we're going to we're going to play out some some intelligence explosion scenarios today okay so let me start with what's the premise here like when we're talking about an intelligence explosion what is actually the premise so um the premise that we're working under is that there is intelligence in text in words and how those words are connected to each other okay so what I have here is a this is a dish rag you imagine that it's filled with the collective intelligence in all written human human text and it's filled with this kind of water and so somehow with deep learning and the Transformer architecture we're able to squeeze out and mimic that intelligence that exists inside that dishrag so you can imagine that all the written data in the world is like a dishrag and and and what deep learning is able to do is squeeze the intelligence out of it and the reason the reason that this works at all uh on a basic level and we've known this since the 80s the late ' 80s uh well not we a couple of people has known this since the late 80s and the early 90s that uh there's this foundational Theory called the univers the universal approximation theorem uh which describes that a neural net given enough layers and the right weights can approximate any mathematical function so when you hear that AI is all that it is an a an AI model all that it is is a file with different probabilities and words remember that that is enough at least in theory to approach any cognitive abilities that can be codified and embed any underlying World model so the theory there is solid mathematical theory for this we're not just you know making stuff up I'm sorry I realized now that I have the same chart that Brian Johnson just showed you um what I want to talk about here is how much intelligence can we actually squeeze out from text so we probably cannot squeeze out more intelligence from text uh than what is in very well written smart human literature the old adage of garbage in garbage out probably still holds true even if you had you know the largest neural net trained on a billion dollar AI cluster if it's all if the training data is filled with zeros that's not going to create a smart model so how smart can that data get well so let's start with the premise that if you go to if you are able to squeeze out the intelligence from the smartest books how smart these models could get is at the level of a very very smart human and we have some some evidence that that um supports this that um if you we have some evidence that if you simply throw more compute at the problem of trying to train a better model there's a limit uh where simply spending more computational resources in training does not minimize does not further minimize validation loss um but uh and this is important the premise isn't that we're just going to throw more gpus at this and train llms and we'll reach super intelligence the premise is that we'll will'll be able to train neural Nets to at least get somewhat close to being as smart as really really smart humans and that's when they'll be able to start to recursively self-improve by doing the research to figure out how to further un hobble itself how to algorithmically improve its efficiency and explore other architectural changes uh like using slow and fast weights for example uh the end goal basically being what what Alpha zero did for chess being able to play against itself until it had enough synthetic data to beat any other computer any other other human in in chess um but we don't need to build that we just need to build an AI that is able to self uh to recursively self-improve itself and and we're seeing some of that stuff um happening today so we I I guess some of you people have maybe tried uh the latest model uh 01 preview from open AI uh so they're able to improve the capability of the model simply by allowing gp4 to think longer this is called a variable inference times it's it it does inference for varying lengths and the plan is to use the The Chain of Thought outputs given that they can be graded for reinforcement learning as synthetic data to the next model just to be clear the 01 uh preview model is not a new base model it's not GPT 5 but the data that's generated from that in those trail of thoughts that it produces can be used to drain something more powerful than GPT for so we're already seeing some of this sort of recursive self-improvement effects although human Engineers are still uh had in charge of that so let's talk about what this means for crypto uh because most of us like I see we're running around doing the same stuff we're investing in the same kind of crypto companies we're thinking the same way but the world is not anymore the way it used to be so let's talk talk about what's going to happen if we're actually in an intelligence explosion which most of you seem to think that we are so there's going to be a fight for attention they're making super intelligence we're making ass points okay so that leads to competition for Capital there's going to be a fight for Capital and S deep I love you but what I the the announcement of of of sentient I have no idea idea what you guys were talking about but we're seeing like $85 millionar uh seed rounds in AI uh and that is becoming sort of less common in crypto so we're seeing increase the hot ball of money that you talk about is used to be in crypto it's starting to shift to be more in AI because again we're making ass coins they're making super Talent JS so it's not weird that this is happening and when you think about it like if you're a VC and you invest in a crypto project usually you have a three-year time span until you get liquidity on that investment so traditionally you give a team some dollars to go work on some crypto infra for 3 years parallelized evm intense based systems and then there's an intelligence explosion maybe now the machines are better at writing parallelized EVMS than than we are so you basically have no idea what the world looks like on the other side of that investment and so the whole VC Playbook that we're sort of trudging along with does it even work anymore so what happens what happens now do we all just simply flee crypto like there's nothing here and so I would be completely comfortable happy to say if it was the case that I believe that crypto has no relevance in this future of an AGI uh an intelligence explosion but I've I've thought about it for many months and this is the best I'm coming up with and I think what I'm going to what I'm going to show here is that there is a role for crypto to play in this environment and I I did not uh the this twirling animation here that's that's not intentional it was the only animation that for some reason keynote allowed me to do so if there's a if there's an intelligence explosion that will lead to a global a GI arms race will be the most powerful technology that a nation state can use against other countries to do cyber warfare on each other to build to control armies against each other there will there will be an AGI arms race and if there's an AGI arms race well what does what what does that mean so in during World War II the United Kingdom and Japan borrowed over 100% of their GDP the United States borrowed 60% of their GDP which would be $17 trillion today and same thing World War I uh United Kingdom France Germany all of them bored 100% of their their GDP so we would have massive Fiat Fiat printing if we went into an AI arms race um but also have geopolitical termoil like we would start to compete for these resources that for example you know China is next to Taiwan Taiwan has tsmc tsmc makes the chips that goes into uh Nvidia that create that that creates the chips that allows us to train uh more intelligence models um so there's it's unlikely to see how the world would be stable in an environment where all these different nations are competing for the control of AGI and then trade becomes more difficult because the currencies that we're using they're inflating ter the geopolitics is more um you know infected and convoluted than it's ever been before and so how are we going to do Commerce when sort of the trust between countries are broken when everyone is sort of in a cold war between each other can we use if we can't use Fiat currencies we don't trust each other's Fiat currencies well can we use oil can we use can we use gold well not if that gold is in a specific vault in a specific country so maybe uh maybe Bitcoin becomes the only way that we can transact with in in this kind of world maybe Bitcoin is the only neutral um nonphysical instrument of money that we can use so then um if that's the case then we really got to fix Bitcoin we get we really got to fix lightning and the scaling issues we don't have more time so if you know what opcat is it's an upgrade to bitcoin this is something that if if you want Bitcoin to be able to sort of be a part in this AGI arms race and act as a currency then then support opcat but that's a different discussion um but we should also think about so okay so how will Bitcoin mining compete if there are AI superclusters that dominate the energy grids let's say an AGI arms race your average AI supercluster is connected to the largest energy arteries on Earth uh as they would be if it's a matter of like absolute National Security and let's say that these on average these super clutch clusters are connected to power sources that consume 100 times the capacity of all the Bitcoin mining farms in the world would this pose a threat to the Bitcoin Network that is only consuming a small fraction of the of the electricity and there's data centers that has access to 100 times that well you may say no but then let's talk about the uh the as6 chip Farms so let's say that you're not worried about the energy asymmetry and but how about then if you consider the fact that today $500 drones are destroying $5 million tanks in Ukraine so let me add in some drone swarms here in this picture um so in a world where Bitcoin becomes the main instrument for trade during the AGI arms race are you sleeping fine knowing that our as6 chip Farms are centralized in single locations and can basically be be destroyed by say a form of drones where each drone costs 500 bucks so this leads you into an interesting conversation so so maybe proof of stake based coins Fair better I think someone at this conference has to come up with a with a a bull thesis for ethereum if you guys have seen the the eth BTC chart lately um so this is kind of a dream scenario for coiners right for the coiners in the room uh a lot of a lot of shitcoins are based on proof of stake so maybe you guys still have a chance uh except for the fact that you know if if drone you know this would only be the case if if drone swarms are all over the place and the world is basically in dis array and in chaos at that point but at least your proof of stake coins might have some utility um now so this is perhaps what people want to know about the most isn't crypto in AI a marriage made in heaven what if we can subdue super intelligence using decentralization so that no evil government no centralized entities would wield the keys to this power no um so given that we don't have uh a way to efficiently verify AI computation you can't use the centralized mechanisms to to incentivize uh inference or training and even if we could uh the global Grassroots Community is not going to be able to amass the kinds of AI h100 chips necessary to be a serious rival in the AGI arms race especially not in a global arms race where the supply chains for um chip manufacturing is heavily controlled but uh AI will need will need crypto for that sweet sweet fast fast exit liquidity because if you think about it if if you are uh a person who building an AI company how are you going to give your investors an exit path on that liquidity if the whole world basically is you know changes in three years so how how how are you going to invest in an AI project if the you know the traditional Capital formation process doesn't work if the whole world changes in three years so how are you going to get liquidity on your AI investment well basically AI projects will have to marry crypto coins so that they can get access to liquidity that they need um I'll make one caveat to that and um that's this is um I won't have time to to to fully explain this but um there is a concept called futarchy and it's a forward-looking way for companies and Nations to self-regulate using prediction markets and by policing themselves uh in favor of outcomes that would be the most profitable so one thing that we have been successful building in crypto are prediction markets prediction markets is the is a Cornerstone to futar key and stable coins are a Cornerstone to prediction market so if you wanted to do something in the crypto AI field don't go the don't try to like decentralize and win the AGI arms you're not going to be successful at doing that but you can leverage the powerful incentivization mechanisms that we have in crypto the prediction markets to allow inform uh for example open AI to give them more data points on how powerful is their next model going to be and allow people to bet on that and create the financial incentives for people to self-regulate so this is a bigger topic we're going to go into something a little bit more simple if this is like way out of your head so one thing that you can do um is you can buy a bunker so bunkers are Maybe the most underpriced real estate in the world right now you can buy a 60s 70s military bunker um for a fraction I'm talking single digit percent of what their production cost is so when if if people start buying up bunkers um and you can buy one now for let's say $800,000 and the production cost is 20 million well you know that's kind of an underpriced asset and at least then you know you'll be able to keep your employees safe from drone drone swarms so that you can keep your crypto projects going that's uh that's actually that's uh that's it for me thank you very much my name is uh Eric Wall by the way I'm ERC WL on Twitter thank you guys