Okay, I think we need to set the gas on this thing higher. So okay, perfect. So I think the most important fact about the crypto space this year is basically that this year is really the year where we turn the corner on crypto and the underlying technology getting to the point where we can actually do all of the things on top of it that we have wanted to do the whole time. So I think if you think back to 2013, who remembers some of the early Bitcoin stuff? Who remembers back in 2013 when we had stores, people trying to get physical stores like restaurants trying to accept Bitcoin for payment? So okay, great story time. Here on the left, this is a bar in Berlin that was called Room 77. It's in a district called Kreuzberg, and there was an entrepreneur named Jörg Platzer who runs this bar, and he decided to make a project out of going around through all of the restaurants and stores in his neighborhood and trying to convince as many of them as he could to start accepting Bitcoin. And he actually succeeded. He got at least 11 places, I think quite a bit more, all within about 500 meters of each other. There was a small region of Berlin where you could just go and you could do basically everything with Bitcoin. And this actually kept going for quite a few years. Over here on the right, this is a store in Boston called Fenway Smash, and very similar. The owners themselves decided, hey, Bitcoin is cool, let's try accepting Bitcoin as a means of payment. Back then there was this really deep excitement about Bitcoin becoming the great next world currency, and if this is the new money, let's actually try hard to get people to adopt it.
So in 2021, I remember visiting Argentina for the first time. And in Argentina, I remember one of the first things that I noticed is that just about everyone there is using crypto. And they were not just using crypto for open source libertarian ideological reasons. They were using crypto because they really need it. The local currency on average has been devaluing against the dollar for the last five years by about a factor of two every year. The country has been averaging like 30% inflation, and that's an average across literally the entire last century. And people just need some form of savings, people need some form of stable value to send, to use for regular business transactions. And one of the things though that I noticed here is that they were using crypto, but they were not using the blockchain. So I actually made a point of paying people with transactions on chain. So this was when I was walking around on Christmas Day and the first coffee shop I found that was open, I walked in, the owner recognized me, and I asked if I could pay in Ether, and she said that I could. So she had a Binance wallet, and the thing they're used to is Binance because it's free for them. I decided, well, I'm from the Ethereum Foundation, I need to actually use the Ethereum chain. And so I paid, the transaction fee was about like somewhere between $2 and $5, and the transaction took about five minutes to confirm. So a similar thing ended up happening to all of these Bitcoin accepting stores. The fees of Bitcoin were extremely low back in 2013. And if you look at how Bitcoin got advertised in 2013, people were very insistent that one of the most important things about Bitcoin is that its transactions are cheaper than the banks, the credit card companies, and PayPal. They screw you over, they charge 2.9% plus $0.29 on a transaction, and then the fees are three times higher if it's international. Western Union charges like 10%. Crypto is basically free. Well, what happened? Transaction fees went up to like more than $10, more than $50. At one point, the highest transaction fee I ever paid was actually $800. And this was one transaction that took a lot of gas. It was somewhere between 500,000 and a million gas because it was a privacy preserving transfer, but still, $800. So fees got too expensive. And the reason why fees got too expensive is because the chain had a fundamentally limited capacity. So people wanted to do, for example, Twitter on the blockchain. And let's think about this question: what would it take to do Twitter on the blockchain? Does anyone here want to estimate how many times they've tweeted in the last week? Anyone show a number? Okay, 100 is a big number. Any other number? More than 20. Okay, so we'll average and we'll say 60. So 60 times a week, 60 divided by 7 is 8.6. So 8.6 per day. Let's say you wanted to make Twitter on the blockchain with 100 million users. 100 million users becomes 860 million tweets per day. Conveniently, there are 86,400 seconds in a day. And so if you divide, to do Twitter on the blockchain you basically need 10,000 TPS. How many TPS does Ethereum do? It does somewhere between 10 and 50 TPS, at least on the L1, depending on the complexity of the transaction. If it's just Twitter, the transactions will be simple, so probably 50. But still, you have 50, you need 10,000. So any time any application even started to be successful, often that application by itself would cause Ethereum fees to go up and the whole thing would break. This is what happened with CryptoKitties. That was then. Now let's think about what's happening now. This is a chart of transaction fees in 2024 on layer twos. So layer twos are the primary scaling solution for Ethereum. They are networks that are not directly on Ethereum, but they are attached to Ethereum, connected to Ethereum in a way that allows them to inherit the security and all the other important properties of Ethereum, but at the same time they do not require every Ethereum node to actually do all of the computations. And because of that, they can be much more scalable. Rollups are the main type of layer two network, and rollups are limited by the fact that even though computation happens off-chain, a significant amount of data still has to be put on chain. And so they are limited by the amount of data. Now in March of this year, we did a fork called Dencun, and in Dencun we added this new feature called blobs, which adds 384 kilobytes every 12 seconds of extra data space for rollups to use. Now this extra data space increased the scalability of rollups and it allowed the cost of transactions to decrease from between 10 and 50 cents to under one cent. So who here would be willing to use decentralized Twitter if it cost you 50 cents every time you have to tweet? No hands. Would you use decentralized Twitter if every time you had to tweet you had to pay half a cent? Okay, more hands. I counted about six hands. If you take the whole population, this looks like somewhere around 500 people. Multiply by 1.2%, extrapolate to the whole world, 8 million times 1.2% is 96 million users. So great, we win. So even still, obviously getting fees down is not the only problem we have to solve. We have to solve go-to-market, we have to give people an actual reason to use decentralized Twitter. But if the fee problem is not solved, none of the rest of it can be solved. And the big thing that happened in 2024 is that the fee problem actually got solved. That's not the only thing that happened. Another thing is let's look at transaction waiting times. When you send a transaction, how long does it take for the transaction to be included on chain? Bitcoin, a block every 10 minutes, often you have to wait 60 minutes or longer. Ethereum, block time with proof of work theoretically was 13 seconds, but in practice because of the inefficient fee market mechanism, often as a user you have to wait anywhere between 15 seconds and maybe 10 minutes. Now Ethereum post EIP-1559, which greatly increased the efficiency of the transaction fee market, and post merge, which moved from proof of work to proof of stake, the waiting time actually went down a lot. If you use Ethereum today, you only actually have to wait usually about 5 to 15 seconds for a transaction to be accepted. And so back then when I was first visiting Argentina, actually EIP-1559 had come out but most wallets had not yet actually integrated it. And so when you pay for a coffee, you actually have to often wait five minutes. Today you do a similar thing and the waiting period is actually not more inconvenient than the waiting period when you pay with a credit card. So I would say the waiting time problem is basically close to solved. Now we can also look at general application user experience. So on the left you have decentralized Twitter in 2015, on the right decentralized Twitter in 2024. Massive improvement in quality of user interface. So the point of all of these things put together is basically that if you think about what would it take for an application using a blockchain to succeed, then no matter how well you do on every other front put together, no matter how well you do on product market fit, no matter how well you do on marketing, no matter how well you do on getting crazy celebrity influencers to show your thing, no matter how well you do on just actually being useful, if you're stuck on a blockchain which is bad at TPS and which is incredibly hard to use, you're going to lose. This is the single big factor that basically killed applications up until this point. This year, that factor is gone. This year the gates are opened. So welcome to the future. But now we have to actually build things. So I think the first 10 years were focused on the basic technology and making sure that the actual infrastructure actually had the capability to do all of the things that we want to do on top of the blockchain. Now we basically have that capability, and the focus needs to switch to what do we actually use that capability to do? What kinds of things are we actually going to build? What actually uniquely makes sense to build on top of a blockchain?
So I think basic payments and finance continues to be an important use case. Basic payments was the first app, payments continues to be one of the most important apps. International payments especially. I use crypto to do international payments all the time when it's more convenient than doing bank transfers. And there's a lot of people around the world who often just don't even have the kinds of access to basic finance that we do. So continuing to improve finance, including the ability to save and grow your wealth and including the ability to just pay people and actually use it, I think these are both things that are just going to continue to become an important category of applications. And these are the things that are already succeeding today. There are entire countries where more than 10% of the population holds cryptocurrency. Often these are countries where the political environment is the most unstable. If people were using cryptocurrency just to gamble, then it doesn't matter what the political environment is. But good currencies are stable, bad currencies are not stable. The common pattern is that the countries that become the most famous for having very high adoption, they don't have a stable internal financial system, they don't have a stable internal economic system, they're not well connected to global infrastructure. They're places where people really actually need an alternative, and blockchains are proving to be actually good at providing it. Now historically, of course, this stuff has been happening, but it's been happening through centralized wallets and centralized exchanges. The problem with centralized exchanges, of course, is they can be hacked, their owner can turn out to be a crazy guy that gambles all your money away, all kinds of things can happen. So blockchains are a much better solution, and decentralization actually does make things safer. The challenge is can we make decentralization practical? And with the technology of 2024, we can. So this is a wallet called Drift. It's basically deliberately intended to be competitive with Venmo and useful with the same quality and standard of user experience that Venmo has, but under the hood it's using stablecoins for transactions. So I expect that we're going to see many more things that look like this. Decentralized social media, decentralized Twitters. We have Farcaster, it's a decentralized Twitter. We also have Lens. So actually two big decentralized Twitters on top of Ethereum, more and more people actually using them now. Prediction markets are another one of those interesting areas. A lot of people have been following prediction market stuff for a long time, and a lot of intellectuals have been very interested in them. Robin Hanson from George Mason University has been talking about them as an idea for a long time, and he came up with this idea of futarchy, which basically says to use prediction markets as a form of governance. So basically, you can make conditional prediction markets that basically say, well, if we do this, then predict what will happen. If we do something else, then predict what will happen. And you actually look at those predictions, and then based on which predictions you like more, you actually make that decision. Now all of these ideas have been very theoretical for a long time, and they just have not had sophisticated participants. They have not had market volume. You need market volume to attract sophisticated participants, and historically they haven't performed too well. This year I think has been a year of really massive growth for them. Polymarket in particular has just grown massively this year, both in terms of volume and, at least from my own observation, in terms of the quality of participation on it. So what's interesting about prediction markets to me? To me, prediction markets are basically a different kind of news service. On the internet, anyone can go and say things, and people will just go and say all kinds of things in every single direction, and it's hard to tell if someone says something, does this actually matter? Should I actually listen to them? Is it actually meaningful? And I think prediction markets actually can be part of the solution here. So here on the left, this was a couple of months ago. There was a tweet by Elon Musk basically in response to people complaining about UK immigration, and Elon replied and he said, "Civil war is inevitable." Very soon, a prediction market popped up on Polymarket basically asking, "What is the probability of a civil war in the UK in 2024?" The market said 4%. So then I actually thought 4% is too high, so I actually bet against it and I put in that myself. And so I think I'm responsible for like one of those declines somewhere in there. And then now it's down to 3%, and I think now it's even down to 2%. So basically, anyone can say things, but the situation becomes very different when you actually have to put your money on the line. And over time, I've noticed that even mainstream political intellectuals, even the types of people who think crypto is useless, crypto makes no sense, there's no point for decentralization, even they end up retweeting screenshots of Polymarket now. To me, this is only the first example. Just saying what will happen and telling people to make bets, this is like the simplest possible use case. I actually think most of the value comes from doing more interesting and more complicated things. So we can do things like conditional prediction markets. Basically, you can have a prediction market that says, well, what decision will some organization make? And then if they make this decision, then what will some particular outcome be? And you can then look at those probabilities and then you can start to use prediction markets as a way to see what might the consequences of certain decisions be. And there are actually already existing examples of some specific markets on Polymarket basically already being usable for this purpose. Another category that's interesting is micro prediction markets. So if someone wants to ask a question, any question, at a very low budget, and this could be any question. One particular example that I'm interested in is basically very low cost dispute resolution. So if you have any kind of subjective question, for example, let's say someone made a tweet and all you want to ask is, is a statement in this tweet true or false? A group of humans could make a reasonably good answer for this, but a group of humans takes a lot of resources to bring together. You have to pay them, they have limited time, and so you cannot do this for every tweet. But here's what you can do instead. What you can do is you can make a prediction that says, well, if we did bring a group of humans together to decide, then how would they answer? And what you can do is you can make this prediction market, and then instead of humans playing on it, you can have AI playing on it. You can have LLMs that just play, and anyone would be able to join the market, anyone would be able to participate. And then basically you create a game where over time the best LLMs would be the ones whose predictions actually do match what the humans decide in the randomly selected portion of the time when humans actually get in. So basically, any kind of mechanism that is high accuracy but which is slow and expensive, you can basically use a prediction market and have LLMs play in the prediction market. You can make a kind of distilled but much more efficient projection, a simulation of that mechanism that is going to be good enough for at least some percentage of tasks. So I think this is one of those use cases that's interesting.
So this gets into this other big question, which is what role should AI play? I think lots of people talk about AI and blockchain intersections. And I feel like there's a lot of people who understand this question on this very superficial level. To me, level zero is basically AI is cool, blockchains are cool, therefore AI plus blockchain is cool squared. I think you can go one layer higher and you can say things like AI is powerful but it's the technology of centralization, it has monopolies, it's not transparent. Blockchains are the technology of decentralization. And so maybe if you can combine them together somehow, you get the best of both worlds. And a lot of people do make this point. It's a very sort of high level thematic kind of point. The challenge is how do you take that high level point and actually translate it into a specific concrete thing that actually uses AI and blockchain in a specific way that actually solves some specific weakness of using only one or only using the other? And I think a lot of answers to this don't make sense, but I think there are a few answers that do. I wrote a blog post at the start of the year where I outlined four categories of AI and blockchain interactions. And I call them AI as a player in the game. This is basically where you have a blockchain smart contract application that defines the rules of the game, and then AIs play within the rules. AI markets, I think, are a great example of this. One example of this that's actually been around for over 10 years is arbitrage bots trading on decentralized exchanges. That is one AI-crypto intersection that has existed basically since at least 2015. So that's AI as a player in the game. AI has an interface to the game. So you can have AIs that help users interact with the blockchain world. You can have AIs that help users understand what are the things that they are interacting with. If they perform a particular action, what are the consequences going to be? What are the risks going to be? What might the security risks be? Then you might have AI as the rules of the game, actually creating an application where the AI actually plays some kind of direct role in top level decision making, like being in a DAO or something similar. The rules of the game, I think, is something that's fundamentally much more dangerous. The reason basically is that if an AI model is open and transparent, or if a model is closed and opaque, then you have to trust it. You have no idea whether or not it's actually being trained correctly. You have no idea if it's being run correctly. If it's open and transparent, you have the adversarial attack problem. People can download the model, people can use gradient descent and they can locally train optimal attacks to try to trick the model into doing whatever they want. Someone even experimented a couple of years ago where they managed to optimize a mask, a physical mask that you can wear on your face, where if you go around a certain type of security camera, it will either think that you're not a person or it will think that you are some different person. So adversarial attacks are scary. And so I think if you want to do AI as the rules of the game, there are a lot of these big challenges. And I think in this case, not blockchains but cryptography, in particular ZK-SNARKs and that whole category of things, can help. So basically, you can have the inference process, maybe the training process, have stronger properties in terms of verifiability and trustlessness than you can otherwise. And one of the interesting things about AI and cryptography is that neural network inference is computationally expensive, but almost all of the computation is matrix multiplication. Matrix multiplication is linear, and linear operations are generally much, much faster and have much lower overhead to accelerate or to do cryptographically than any other type of computation. For FFTs, it's less than a factor of six. Now someone has managed to ZK an entire AI LLM computation where the overhead is, I believe, about 110x. Now 110 is a lot, but if you compare to any other computation, ZK overhead for most regular computation is going to be like 10,000x. So already a pretty massive improvement here. So potentially there is an opportunity to try to get the overhead of the cryptography actually low enough to the point where you actually can do AI computations cryptographically in such a way that you can do things like decentralizing models, preserving privacy of data, preserving privacy of the model, getting integrity of the training process, in a way that is low enough that it actually is competitive compared to just basically making a much simpler model and just running it locally on your laptop. So I think this set of areas is also interesting.
Now one last area that I think is interesting to talk about is a return to enterprise applications. So who here remembers the era of permissioned consortium chains from like 2014 to 2019? Okay, a lot of people remember it. So the core insight behind the attempts to do blockchains in enterprise is basically that you have this technology which is very decentralized, but it has scalability problems, it has potentially privacy and regulatory problems. It has these problems. And you have existing workflows that are centralized but they're too centralized, and so you have silos. Maybe a compromise between centralization and decentralization, you create some kind of structure that's somewhere in the middle, then you can make something that actually really works for large scale and regulated enterprise applications. The problem is that I think permissioned blockchains in particular ended up being the wrong compromise. Basically, I think they ended up being too centralized to benefit from decentralization, and too decentralized that you don't get the convenience of centralization at the same time. Basically, as soon as you go from one node to multiple nodes, you have to worry about networking, you have to worry about PBFT, you have to worry about all kinds of asynchronous stuff, latency, you have to worry about a whole bunch of different things. But at the same time, if you have only like three nodes or 30 nodes in the network, then the problem is that when your consortium starts, people who come in might trust it, but then it turns out that it's hard to scale it further because once you scale it further, it just gets perceived as being something that's controlled by whoever the first group is. So I think now we have ZK-rollups, and in particular we have ZK layer twos, especially validiums. And I think we now have an opportunity to actually go back to enterprise blockchain applications but try them a different way. Instead of creating permissioned chains, we embrace the public chain, but we embrace it as a base layer. We do not put all of the data on the public chain. Instead, we publish only the hashes, only the roots of the data onto the public chain. Everything else can still happen even in a server. You can keep your server, you can keep the exact server based code that you have. You just add a little bit extra on top to publish hashes to the blockchain and to publish proofs to the blockchain proving that whatever updates you're doing, you're actually doing them correctly and you're following the rules that you said that you're following. So if we do this, I think this is actually a compromise between centralized and decentralized approaches that might actually work much better and be much more viable for a lot of different applications. So summary: basically, I think the technological fundamentals of the crypto space are brighter than ever. So many things have massively improved. The difference between using the ecosystem today compared to two years ago or five years ago is huge. There is, I think, lots of consumer and public demand for the underlying ideals of building global, open, and trustworthy systems. There are lots of people stuck in areas where financial systems just don't work well. There are lots of people who want alternatives to Twitter. There are lots of people who want alternatives to centralized identity systems. There are people who want some kind of more open and more decentralized alternative to the way that AI is turning out right now. So there's a lot of demand for these things. The challenge is that crypto the set of ideals has always met this demand, and crypto the set of technology historically has not been strong enough. And we need to actually continue improving the underlying technology, and we need to actually build these things because the base layer actually is enough now. But if we want these applications to succeed, then the entire stack needs to be built. Thank you.
