Web3 + AI at ETHDenver | Part 2

Hello, everyone, and warm welcome to all newly joined subscribers!

This is the second portion of the most compelling ETHDenver talks on Web3 + AI topics I summarized for you. As I said in Part 1, the 2024 edition of the world's largest crypto event gathered so many Web3 + AI builders and experts that some people started calling it AIDenver instead. Since so many of us could not be there in person, I decided to collect the best recordings to watch!

Thank you for being here! Let's dive in!

AI Agents With Crypto Wallets: That's a Good Thing... Right?

This panel discussion moderated by the former Chief Content Officer at CoinDesk, Michael Casey, brings three perspectives, which in some other context would have been conflicting, but here are more or less in sync. CoinFund's Jake Brukhman represents the investor point of view, Niraj Pant talks about building the Web3 + AI project Ritual, and Stacey-Ann Pearson participates on behalf of Amazon's AWS, but in her role as their Head of Web3.

Each one of the speakers shares the Web3 + AI use case they are most excited about, from using crypto as a tool for self-sovereignty to ensuring value redistribution within AI development.

Brukhman states that reinforcement learning models are already much better at finding vulnerabilities than humans, since they discover them not only at smart contracts, but also at the VM itself, something that standard auditing can't do.

Pearson, on the other hand, declares that Amazon doesn't bet on any particular foundation model to win the AI race. It rather focuses on providing infrastructure and access to as many models as possible. On the Web3 side of things, Amazon is excited about the IoT element of the Web3 + AI convergence, whereas the company looks into use cases and applications involving agents communicating directly with a crypto wallet.

Listen to the full conversation to learn why the big monolithic models coming from Google, Microsoft, and OpenAI are not going to be the answer the world needs.

On-Chain AI Inference: Securing & Scaling Blockchain 

Jake Brukhman again, founder and CEO of the investment firm CoinFund presented his vision for the intersection of Web3 and AI, and what motivated him to start investing in it as early as 2021. He distinguishes two categories of Web3 + AI applications and lists the most interesting use cases currently being developed in each one of them:

Bringing AI to Web3:

• LLMs as on-chain assistants
• DeFiMl for modeling risk, predicting volatility, and optimizing yields
• Reinforcement learning for security

Bringing Web3 to AI:

• Introducing Web3 primitives into the pipeline producing AI, aimed at democratizing resources like data, compute, and capital
• Open marketplaces for data, models, inference
• Verifiable compute using techniques like ZKML, and opML
• Self-sovereign data and personal agents
• Token incentives for AI models optimization and fine-tuning

Brukhman outlines the latter category as much more important, especially since today's AI pipeline is almost entirely owned by a very small number of big tech companies. At the same time, though, most of the innovation in AI is actually happening in open-souce, and the best way to keep it safe is to make it as open and transparent as possible.

Listen to the entire talk below to learn more about CoinFund's investments in Web3 + AI companies, including Worldcoin, Bagel, Polywrap, Allora, Gensyn, and Giza.

MPC, ZKP & FHE: Which Cryptography to Bet On (Hint: All) 

Starting with a bit more technical, but equally interesting talks. Guy Itzhaki, CEO of Fhenix, gives a long-form explanation of the different data confidentiality solutions currently in use. But first, he argues that blockchains' transparency and lack of confidentiality is not always a good thing. Rather, it could be a huge problem and a limiting factor to our ability to create new use cases that would actually grow the crypto ecosystem. Here is his guide to MPC, ZKP, and FHE:

Zero-Knowledge Proof (ZKP): It is largely used in rollup validation, but when talking about using it for data confidentiality, it is quite limited in what it can offer. Since it is a proving mechanism, which does not enable computation on encrypted data, it can provide confidentiality in very small amount of use cases. For example, for products like auctions or prediction markets, ZKP won't be the right solution.

Multi-Party Computation (MPC): A technology that enables a group of different data owners to jointly compute a function of their private inputs, without revealing their private data. Only when everyone has completed their share of the computation, the total computation can be executed. MPC in the blockchain domain is used mostly in custody wallets. It requires significant communication, so its use cases are limited to a small set of participants.

Trusted Execution Environments (TEEs): A secure, isolated environment within a CPU that allows you to do computation securely, protecting it from external interference. It provides a good performance, but its security guarantees are lower than a cryptography guarantee.

Fully Homomorphic Encryption (FHE): It enables computation on data while it is encrypted and is very easy to use. Fhenix, the first-ever FHE-powered L2, is built on the Arbitrum stack, based on Solidity and uses a Zama FHE library.

I will leave to you to listen to the full presentation to understand why we need all three solutions at once!

Bring On-Chain Confidentiality With Fully Homomorphic Encryption (FHE)

One more FHE talk, this time delivered by Remi Gai from Inco Network. Inco is a modular confidential computing network, based on Cosmos. Again, Gai starts off by confirming why we need to have confidentiality in blockchains, but goes on to claim that FHE is a much better solution than ZKP.

Since FHE allows for computing encrypted data, when applied to blockchains, it is able to ensure both the privacy and composability of on-chain data. When applied to an EVM, called fhEVM, the result is:

• Encrypted state stored fully on-chain
• Computable encrypted state, without decrypting
• No TEE required, nor circuits (as opposed to ZKP) or centralized off-chain coordination
• Native (on-chain) randomness (very useful for games like poker)
• No actor coordination (MPC), since the blockchain is the coordination layer

Inco targets to become the universal confidentiality layer of Web3 and to cater to various other blockchains, since integrating FHE comes with quite high hardware requirements.

Potential use cases of fhEVM, which would not be possible on public chains, include gaming (like private betting or games like Mafia), DeFi (use cases like secret auction), enterprise solutions (private payroll), and others like private voting.

How AI Based Threat Detection Can Save Web3

Andrew Beal from Forta, a monitoring company for L1 and L2 blockchains, deliberates on employing AI for blockchain security, and using machine learning to detect malicious actors.

Beal describes the four stages of an exploit: funding; using funds to deploy a new contract; invoking the contract or what is considered the attack transaction; and laundering the stolen funds. Forta's previous tactic consisted in threat detection, but now they focus on threat prevention. This is why their goal is to detect the attacker's contract deployment phase and flag it then and there.

Beal also explains that Forta trains a model to tell the difference between a malicious and a normal smart contract. They have 4-5 years of malicious smart contract data, a lot more of benign contracts data.

Thanks to the fact that the detection companies and tools are much numerous and more sophisticated today compared with a few years ago, most smart contract exploits are detected in advance. However, that doesn't necessarily mean that they are prevented. Hopefully, AI will help on both those fronts.

Disclaimer: None of this should or could be considered financial advice. You should not take my words for granted, rather, do your own research (DYOR) and share your thoughts to create a fruitful discussion.