Balaji: Eight Use Cases Interpreting How Crypto Rebuilds Trust in the AI Era?

Video Title: AI Makes Everything Easy to Fake---Crypto Makes it Hard Again | Balaji Srinivasan at SmartCon 2023

Video Author: Balaji Srinivasan

Compiled by: Qianwen, ChainCatcher

About the Author:

Balaji Srinivasan, angel investor, technology founder, and author of the Wall Street Journal bestseller "The Network State." He was the Chief Technology Officer of Coinbase and a general partner at A16z, as well as an early investor in many successful tech companies and crypto protocols.

Recommended Reading: 《Balaji Srinivasan: The Genius Investor with the Most Moves in Crypto》

The following is the video content:

Today, I want to talk to you about two well-worn topics: artificial intelligence and cryptocurrency—specifically, how AI makes everything easy to fake, while crypto technology makes everything real again.

There is a specific intersection between the two, that is, generative AI can easily fabricate content online, so how do we verify it? How do we, in a sense, restore the scarcity of information ? This is where cryptocurrency /technology comes into play. As I just mentioned, AI and blockchain often intertwine, and this article will outline eight overlapping areas between them, interpreting how we can use crypto to rebuild trust.

AI makes forgery easy, while crypto makes forgery hard. Here is a photo of Trump being arrested, which is AI-generated. News media might say we can tell whether this is AI-generated content by looking at the fingers, as AI currently cannot accurately reproduce fingers, but these technical issues will eventually be resolved.

So, fundamentally, the most important question is how you want to verify signatures on Ethereum. You want to see that these images and content are digitally signed, preferably through an ENS (or something similar). This way, you can confirm that it is this ENS name, along with the Ethereum address and ENS public key associated with it, that generated this content.

AI generates content, crypto verifies it. In fact, we already have some concrete frameworks for crypto verification, such as ENS/IPFS. If you have the content hash, you can retrieve it and map it to an ENS name. For example, if the content is signed, you can use it to determine whether it is AI-generated or human-generated content. Of course, humans can also sign AI-generated content, but at least you know where it comes from—this ENS name (ENS does not necessarily have to come from an individual; it can also come from a company, etc.).

AI propagates information, crypto verifies information. This is important because once there is a source, once there is a citation, AI will perform crypto verification. For example, there is a service called "perplexity.ai," and last year I asked it about the FTX hacker. Ideally, you want these citations to be on-chain. Some people might think that only certain content can be proven on-chain, like financial records. I agree with this view, but it’s a bit like saying in the 90s or early 2000s, when there wasn’t much internet content or web content, you didn’t know how fast it would develop in the future.

Therefore, one way to think about it is that AI breaks the public network, but Web3 builds a trust network. Now the internet is filled with AI-generated false content; maybe Google can solve this problem, but the ideal solution is still applications like "interface.social."* You will find that* there is a wider variety of data on it, not just financial transaction data, but also social interactions and so on. It actually shows what a good trust network looks like— many interactions are on-chain , and you can cryptographically verify many different aspects of these interactions. It’s not just about verifying a single action, like this entity signing this content; you can also verify all other actions of this entity, allowing you to start calculating whether this entity is human. This is why we can establish Web3 trust. You are not just looking at what this**** *public key* *has signed in transactions, but rather looking at the macro interaction of this* *public key* *with other* *public keys* ****across the entire network.

Our current network has a high weight in determining whether content is "real," such as relying on Google's page ranking to make judgments, but actually, many false contents also rank high, so this method is unreliable. Therefore, we need a decentralized trust network that anyone can view, anyone can index, view on-chain data, and visualize it in these next-generation block explorers, displaying it like a social interface.

AI makes CAPTCHAs ineffective, crypto rebuilds CAPTCHAs. As shown above, a robot is clicking on a CAPTCHA saying "I am not a robot." But cryptocurrency can change this situation; if you sign them with Ethereum, you can require small payments, request to view payment history, or pre-stake some funds, etc.—in other words, raise the cost of forgery. Robots can still log in using Ethereum, but you can charge them high fees to prevent them from doing so.

AI training is usually centralized, while crypto technology can decentralize it. Now we have centralized training and centralized models (like Open AI), and we also have centralized training and decentralized models (like LLama2). But ideally, we also want to adopt decentralized training and decentralized models. If you look around, you will find that most of these projects are cryptocurrency-related. You may disagree with what these projects are doing, but that doesn’t really matter; the point is that we can raise a lot of funds through cryptocurrency crowdfunding, and we can use this funding to train AI models.

When you train these models, you are not just training them in a decentralized or partially decentralized way. What I mean by decentralized is that at least the funding is decentralized. You may still need to train them in a centralized cluster, but at least people are involved.

AI evaluation is centralized, and crypto can decentralize it. Now you can evaluate LLama 2 on a Mac studio. This actually means we can approach a state where everyone training a model can run it on powerful hardware, similar to running a Solana node.

You can imagine that whenever Ethereum or Solana upgrades, people update their nodes, update models, and hold models; perhaps each model evaluation requires spending tokens, so those who fund the models can receive more tokens, and then they can also pay for model evaluations. This is just one way to think about it. But I believe that liberating these things from centralized actors is a worthwhile issue to consider.

AI creates many power centers, and crypto can decentralize them. There has been much discussion about AGI issues. I think an implicit background assumption is that people believe there will be a large unified cult-like AGI (as shown in the image above). But if our community can realize the concepts of decentralized funding and decentralized evaluation, you can imagine a polytheistic AGI (as shown in the image above), where each community has its own better version of an oracle, and they can ask the oracle questions, such as how will the numbers fluctuate? For example, what would George Washington do? Different societies and communities will have their own intelligent systems that they can query. They can ask the intelligent systems about citation situations, and the intelligent systems can even provide on-chain citations.