Thank you. I'm the co-director of the Initiative on Cryptocurrency and Contracts, or IC3, and I'm excited to tell you that today we released a survey on the intersection of cryptocurrency and artificial intelligence, and it's a real wonder. It covers over 150 pages. So in this report I will only be able to briefly introduce you to the article. I am also in the unusual position of unusual position of telling you about an article that probably has more authors than people in the audience today.
You can see them all here. So, as I said, I'm going to talk about the intersection of cryptocurrency and artificial intelligence. They don't really go together naturally, do they? I would venture to say that they are not natural satellites. Why am I saying this?
Well, cryptocurrency is complicated in a way . By this I mean that it is initially purely software-based. For example, if you propose a smart contract to create, say, 10,000 NFTs for, say, 05 ETH, the contract will do just that. It doesn't matter what the market conditions are. It doesn't matter what the wishes of the contract creator are.
If the contract was created specifically for this purpose, then it will accurately execute its instructions programmatically instructions programmatically . Another feature of cryptocurrency that makes it difficult to work with in the sense in which I mean the term is that it is unforgiving. And this is well illustrated by the Kelpdow attack a few months ago. What happened during this attack? It was an attack by the North Korean group Lazarus.
Well, it was a pretty complicated operation. It included a included a denial-of- service attack that caused a failover to a compromised node, etc. But in the end, it was just one corrupted message. And that one message led to the theft of hundreds of millions of dollars. So one message led to Kim Jong-un being able to significantly expand his two favorite collections, his wine collection and, of course, his nuclear warhead collection.
That's what I mean by relentless, right ? A small mistake ? A small mistake , one poisonous message can have serious consequences. As a result, the crypto industry has pioneered the development of a range of security tools security tools , including blockchains themselves and smart blockchains themselves and smart contracts. You can think of them as a security technology, but also as a range of enabling technologies, privacy and integrity tools that include things like trusted execution environments execution environments .
I'll talk a little bit about that today, as well as zero- disclosure proofs, secure multi-party computing, and so on. When we use the term cryptocurrency in the survey, and since I'm going to use it today, I mean all of these things, blockchains and the technologies that power them. Okay, this is cryptocurrency, which, as I said, is complicated. Artificial intelligence is soft in the following sense. We don't really understand how frontier models work, and as a result, their behavior is unpredictable and error-prone.
As a simple example, Chat GPT is really smart. He can pass the bar exam. But on the other hand, you can ask him a simple question like this. Has anyone conducted this experiment? Okay, a few of you, right?
With the correct version of Chat GPT, you can ask it how many rupees are on Raspberry and Coin Chat GPT's answer is two. And I hope we all understand that this is not the right answer. Or you can ask him to, say, draw a rabbit and you'll get something like this . You'll notice that this rabbit has more than a full set of ears. So in this sense, AI is soft.
Well, as I said, it's unclear how these two technologies fit together, but there are examples of their intersection. And I think a very good example is Worldcoin. Okay, so what is WorldCoin? Well, what you see here is a photo of people lining up in Nairobi, Kenya, in 2023, to present their faces to this mysterious and disturbing object known as the orb. They register their irises to create a database, a registry of people on planet Earth.
Well, it's a crypto project, but they recently renamed themselves World, and they've gone through a few missions, but their current mission is to create an identification network that will distinguish between real people and bots. And this is thanks to the spread of AI. So this is a good demonstration of, as I said, the fact that these two technologies can actually be closely related to each other. And in fact, among the co-founders of the world is none other than Sam Allman. Good.
So, how can these two technologies work together? In the survey, we present a model present a model according to which AI can serve as middleware for cryptocurrencies, and cryptocurrency can do the same for AI. For example, an AI adapter can take human expressions of intent or data from the physical world and transform these inputs into inputs suitable for cryptosystems. And cryptocurrency can similarly take input from the physical world or for web servers and protect it, ensuring integrity and confidentiality as input to AI pipelines. And I'll give you examples of both of these types of middleware deployments .
Let's talk about the AI crypto adapter about the AI crypto adapter . By this I mean the ways in which AI can improve cryptocurrency. I'll start with a simple example of a smart contract. This is an example that people like to use to demonstrate how oracles work. This is a contract for airline ticket insurance.
Anyone who is going to fly can request a policy in the contract, pay a small premium so that in the event of a flight cancellation or delay, compensation is automatically paid. Or, in this case, $100 is paid automatically if Alice's flight is delayed or canceled. Well, that's all well and good, and a nice demonstration of how oracles and smart contracts work, but the complexities of cryptocurrency lead to some problematic limitations in this particular example and many others , and in particular lead to smart contracts being only suitable for a narrow range of applications. The type of airline ticket insurance contract I showed you is known as parametric insurance. This is essentially a bet on whether the flight will arrive on time.
And this type of betting is problematic from several perspectives. First, there is the problem of moral hazard. You can take out a policy and then call the insurer for the flight. This is clearly not good. But beyond that, there is no way to document the actual damage suffered by the traveler and compensate him accordingly , which is exactly what insurance is supposed to do, right?
Insurance is designed to compensate for specific damage, which is why parametric airline ticket insurance does not have the full set of real-world insurance features, so to speak. A large insurance company called AXA has been experimenting with parametric airline ticket insurance. They launched a product called Fizzy, and Fizzy no longer exists because of these issues. Good. But let's say we take a typical Oracle, which today just gets data, typically from an API, and aggregates that data or processes it in a very simple programmatic way, and we enhance it with a machine learning model, for example, with LLM.
Then we can do some really interesting things, and we can turn this parametric airline ticket insurance smart contract into something more like a real insurance contract. Specifically, Oracle can receive flight status information, public information, but can also receive private information, such as receipts from a traveler whose flight was delayed and had to stay in a hotel, buy a toothbrush, etc. And the LLM can then determine the extent of the harm suffered by the user and the appropriate level of compensation. And now what we have looks like real-world insurance. And all of this can be done with privacy protection so that the traveler's financial information is not disclosed to the smart smart contract deployer.
So this is an example of how hard and soft elements can actually come together productively. Isn't that right ? This smart ? This smart contract that I just described looks like a real contract, like a real insurance contract. It can use natural language descriptors.
It can handle ambiguity. He can take into account unforeseen circumstances. This is what you want from a real contract. At the same time, the strong cryptographic safeguards cryptographic safeguards in this design I just described ensure that the data comes from real web sources, that these receipts really came from a financial institution, and that the data remains confidential and the user knows which model pipeline is actually deployed in the insurance product. Another application of artificial intelligence in cryptocurrency is artificial intelligence-based investments.
Here, the idea is to use artificial intelligence agents to manage investment portfolios. This is not a crypto-specific idea, but it is an idea that has been actively catalyzed by the crypto industry. For example, a company called AI16Z, which for obvious reasons had to change its name to Eliza OS, does just that and advertises itself as one of the first known venture funds controlled by artificial intelligence. It turns out that this first wave of AI-based investments is mostly hype, as these systems are not autonomous as advertised, and they essentially function as meme coins. They don't have many assets under management.
And this is shown in a recent IC3 article. But the idea persists, the vision persists the vision persists , and people continue to create these systems, and they become more and more sophisticated and viable. We describe these systems or use the term “ collective investment algorithms with coin withdrawal” for these systems. The idea is that investors can raise funds to manage an artificial intelligence system in order to democratize access to the best investment technology— AI-based investment technology. So, graphically, a group of people get together, they send a certain amount of money to an AI strategy , which I'll call S, and that strategy determines how their fund is managed, what purchases and sales they make.
However, in our research we have shown that there is a fundamental and unavoidable paradox in this idea, which we call overselling coins, and it is this: either this strategy is transparent according to the DeFi ethos, but in this case the coin is vulnerable to public frontrunning because everyone knows what trades it is going to execute. So either it is transparent or at least partially opaque, in which case there may be a hidden channel or backdoor, and then the fund will be vulnerable to private frontrunning. In frontrunning. In any case, investors will lose. Therefore, a key research topic for us is an attempt to technically duplicate the protection that trade enjoys through regulations.
Right? We are trying to find guarantees for these AI- based investment funds that do not require human intervention. And all of this is described in another recent article and in a survey. Let's now consider another direction. Crypto AIX: How Cryptocurrency Can Improve AI.
One of the most interesting ideas in this combination of technologies is Dpin, right? Decentralized physical infrastructure networks are a fancy term that isn't really very descriptive. What I'm talking about, to give you a concrete example, when it comes to AI, is the vision that you can create open markets for physical infrastructure. For example, you can rent a GPU for AI, training, and inference tasks, and then people can buy those resources to access them more cheaply, and we can democratize training models at the edge of capabilities so they are no longer controlled by a few monolithic entities. There are a whole bunch of companies that do this and advertise significant savings on GPU rentals .
How real is this technology? The answer is: we don't really know, and we're a little skeptical. The reason is that the most expensive step in training a model is the so- called pre- training, which is performed in parallel on different GPUs and incurs high communication costs. It is very sensitive to network conditions. However, industry proofs of concept do not report how much connectivity is involved in training the model.
So we don't know how well their networks are performing, which is a real question considering they are decentralized and they don't report computing costs. The largest model to date, trained using a pin code, has only 40 billion parameters billion parameters . This is a Consilian model, which is significantly smaller than the modern frontier model. Frontier models have one or two orders of magnitude more parameters, requiring three orders of magnitude more computation. So the bottom line is that we really need more evidence to believe evidence to believe that deep ppin will be viable.
This is for learning. There are other issues related to the findings that were also discussed in the review paper. Now let's go back to this picture I was talking about a little earlier. An oracle enhanced by machine learning. I want to talk about some of the handy things we can do with this design.
One very attractive thing we can do is confidential lending. So, the idea is that we run our oracle and our model in a protected environment, the so- called trusted execution environment execution environment , which ensures correct execution and confidentiality of state data. And let's say Bob wants to borrow money from Alice. Alice is going to borrow money from Bob. Alice can prompt the Oracle and ask, " the Oracle and ask, " Should I give Bob a $5,000 car loan?"
Bob can then provide evidence from his financial institution. He can send this to Oracle in favor of getting a loan. And in a machine learning (ML) model, TE can decide whether Bob is actually creditworthy enough to get a loan. All of this is done without any disclosure of Bob's personal information. Alice the creditor receives a guarantee that the data is reliable.
The decision was made by a specific model. The solution is reliable. Bob knows that his personal information has not been improperly disclosed. Not revealed to Alice at all. Actually, isn't that right?
And there are many other things you can do with this design. You can prove, as a whistleblower, that you have inside information about some important company, such as an artificial intelligence firm, without revealing your identity. You can do more frivolous things. Prove that you visit a lot of Japanese restaurants Japanese restaurants , and therefore you are a qualified reviewer on Reddit. Again , without revealing your identity.
You can also create credentials for the code. You can take your own agent code and prove that it has passed an LLM-based audit without revealing the code. This is especially useful in an agent context, where you don't really know what properties to attribute to the service you are interacting with. We discussed this idea in an article that came out just a few days ago just a few days ago about so-called pyreds. Finally, I want to make a certain warning, which was also addressed in the review article, about what I will call scam agents.
We also call them UAAS also call them UAAS , unstoppable autonomous agents in the review article. And the key point here is that we worry that AI might overstep its capabilities. Money is a vector of escape. Okay, escape. Okay, scam agents with access to cryptocurrency, particularly stablecoins, can use money to survive in the wild, to continue to exist in the wild when they shouldn't.
They can rent servers today with stablecoins from any of many companies. They can use some of the technologies I described , such as tees and blockchain, to create censorship-resistant command and control channels and to obfuscate their code to hide their very existence. And if you want to know what GPT thinks about it in the chat, here is the image he provided. You'll notice the joy on the agent's face, right? Money is also a means of influencing the world.
Using coins, agents can rent people today. You can supposedly do this with stablecoins through rent a human.ai and do things like launch media campaigns and otherwise impact the physical world around us. This is a real concern, and that is what we are covering in the review article. As I said, this is a short overview as it is a very long and extensive article, but we cover many other topics.
Artificial intelligence for blockchain analytics, decentralized data markets, and agent-based sidewalk rails. Most importantly, we list five myths and misconceptions that are prevalent in the industry that are prevalent in the industry , and debunk or clarify them. If you want to know more, as I said more, as I said , the article came out today. You can find it on this website. You can read the entire 150-page article.
You can read the one-page summary or contact the contact the chatbot we created on this page. Uh, you can discuss the article with a chatbot. So, you don't have to read this article at all, but we hope you do, and we hope you find it useful. Thank you very much.
Automatic transcript — names and jargon may be misspelled.