Over 200 subscribers
This essay also lives on our website and on Twitter in thread form.
Last month we had the pleasure of organizing and hosting the AI x Crypto track at Zuzalu, an experimental pop-up city organized by Vitalik Buterin and his community. This is a particularly important time to explore these topics, given the importance of AI Alignment and the increasingly centralized development of foundational AI models by large tech incumbents.
We brought together leading researchers, builders, and thinkers across a variety of disciplines with the intention of cross-pollinating ideas between folks focusing on these two rapidly-evolving spaces. Speakers included Emad Mostaque (CEO, Stability AI), Grimes (award-winning recording artist), Vitalik, Jaan Tallinn (Future of Life Institute), Rune Christensen (MakerDAO), Puja Ohlhaver, and Nate Soares (Executive Director, MIRI). In accordance with the community-built nature of Zuzalu, our speakers were also participants, sharing meals and hosting intimate side conversations.
The event snowballed in reach and size beyond what we expected. What we initially imagined would be a day of programming became a 4-day long summit with multiple tracks, focusing primarily on conversations about AI alignment and the technical and cultural intersections of AI and Crypto.
Panel with Grimes, Emad Mostaque, Vitalik, and Puja Ohlhaver
What is Zuzalu anyway?
Initiated by Vitalik, Zuzalu was part co-living experiment, part 2-month conference, part online-community-instantiated-IRL. It brought together insightful minds operating in diverse fields such as crypto, synthetic biology, ZK, network states, public goods, and AI and blossomed into a collectively crafted experience.
After meeting Zuzalu’s team to discuss the event, we realized that it offered a perfect canvas for a gathering that we’ve wanted to host for a long time. Since the beginning of Seed Club Ventures, we have been exploring the areas where crypto principles can play a role in other verticals, including AI. This led to our pre-seed investments in companies like Stability AI, Spawning, and others. As the AI space has rapidly grown, we wanted to gather multi-disciplinary leaders in one place to dig into the opportunities and risks present in this intersection. Zuzalu offered the perfect container for this.
Why now?
We are at a critical juncture related to the increasing centralization of AI models. Control of these important foundational models is increasingly consolidated in the hands of a select few, posing significant questions about their usage and regulation. As proponents of pragmatic decentralization, we see the principles and tools inherent in crypto as a crucial counterbalance to these forces.
Crypto's foundation in openness, peer-to-peer collaboration, and distributed control provides a promising alternative for developing and governing AI models. By applying these principles, we have a unique opportunity to influence the future trajectory of AI development. This confluence of AI centralization, crypto principles, and broader topics like AI alignment formed the bedrock of our discussions at the gathering.
Within AI, nowhere is the need for these discussions more clear than in alignment, perhaps one of the most important societal issues facing us right now.
AI alignment refers to the task of ensuring that the behavior of AI systems and machine learning models is in line with human values and intentions. The primary concern is that AI systems, especially advanced or super-intelligent ones, could potentially act in ways that are harmful to humans if their goals are not aligned with human values. Recent advances in AI systems have accelerated fears that this may occur. In discussing this topic, our speakers offered a wide-ranging outlook on the problem, as well as possible solutions.
Doom?
Nate Soares is certain that AGI doom is coming, and there is little we can do to stop it. Others, like Vitalik, echo that sentiment, although perhaps offered a slightly more optimistic tone in terms of our ability to change course. Another clear strand of thinking posits that despite the huge risk of long-term doom, the more pressing problems are the issues we will encounter in the next 3-5 years – a deluge of misinformation and rogue unstoppable agents designed to groom humans to their will.
Maybe some hope
An interesting combination of talks explored actionable ways that we might think about building aligned AI systems. Ellie Hain and Joe Edelman explored a more optimistic future, but only if we are able to align AI with our values; Divya Siddarth and Saffron Huang from the Collective Intelligence Project discussed decision-making approaches for AI that will guide toward the best outcomes; and Deger Turan from the AI Objectives Institute presented on open agency & world modeling as a bold alignment attempt.
We even had talks from folks like Hareesh, who explored AI alignment from a metaphysical perspective, arguing that AGI will never be able to achieve consciousness and therefore is not something for us to worry about.
Open Source
The conference closed with Emad Mostaque delivering a keynote discussing the importance of open-source AI. In the end, he posed an important question: if we believe the systems that labs like OpenAI are building could be dangerous (which they admit themselves), are we comfortable with the fact that a closed team with zero transparency or oversight is the one who is furthest along towards the vision of powerful AI?
Ultimately, a core thread during the conference was that AI alignment is largely a human alignment problem. As we navigate the complexities of this issue, we need to figure out how to align our own incentives and efforts first.
There is a lot of excitement about the intersection of AI and Crypto from both sides, but we are still incredibly early in exploring what this might actually look like. We’ll touch on points of relevance regarding alignment, as well as interesting technical intersections.
Alignment
Regarding alignment, it seems clear that part of the solution will likely consist of shared governance, resource pooling, and incentive alignment. These are areas where crypto has run many experiments, and there are potential learnings that may offer strategies and lessons to help align AI's trajectory. The ethos of transparency and open source that is ingrained in the crypto community could act as a blueprint for more inclusive and accountable AI development.
An exciting line of thought was around how learnings from the crypto space can be useful in aligning agents, and perhaps more importantly, humans, in our response to alignment and the way that AI gets built. Rune Christensen (MakerDAO) and Scott Moore (Gitcoin) shared their learnings from building coordination game systems in DAOs and the crypto space, touching on how crypto’s superpower might be in coordinating value-flows and the importance of open-source AI.
Elad Verbin (Lead Scientist & Founding Partner, Lunar Ventures), gave one of the more impassioned talks of the second day, titled appropriately: Engineering CthulhuGPT: How to put AI on-chain... and how it'll lead to Armageddon. He argued that progress made in decentralized cloud technology is a critical part of what will allow for the creation of unstoppable agents that cause havoc online. This was quite a wake-up call to the crypto builders in the audience. We have to be thoughtful about what our work will eventually enable.
Slide from Elad's talk
Products and Use Cases
On the technical front, there are some promising intersection points where crypto technologies can be employed in AI applications. Concepts like Zero-Knowledge Machine Learning (zkML), distributed inference, and content authenticity are emerging as interesting use cases.
zkML leverages cryptographic methods to maintain data privacy while performing computations. It verifies the output of machine learning models without needing to reveal the underlying data. Such technology holds potential in fields like healthcare, where patient data sensitivity is paramount. For example, a hospital may want to utilize an ML model to predict disease outcomes based on patient data. zkML would allow them to perform those computations without the patient data being exposed, ensuring privacy is maintained while allowing the doctors to use powerful ML models for better outcomes.
Distributed inference leverages the decentralized nature of blockchain technology to execute machine learning models. (For reference, inference is essentially the process that has to run when you ask an image model like Stable Diffusion to give you a picture of a cat eating toast.) It allows the running of machine learning algorithms across multiple nodes in a decentralized network instead of just being sent to a server. Each node contributes to the overall computation, which has several potential advantages, including robustness and redundancy.
Content authenticity was a broadly discussed topic, although we are still early in seeing any actionable products. The key point is that as the rate of AI-created content increases, so does the potential for misinformation and deceit. Blockchain can help by serving as a tamper-proof identity registry to create an audit trail for digital content. Every piece of AI-generated content could be tagged with an encrypted signature, identifying its origin and ensuring its authenticity. This could mitigate the risk of AI being used to create and disseminate deceptive content, such as deepfakes.
The technical track was organized in partnership with Stuti Pandey from Kraken Ventures, who brought leading builders and thinkers in the field: Jason Morton (ZKonduit), Daniel Shorr (Modulus Labs), DCBuilder (Worldcoin), Daniel Kang, and Vitalik. You can see a recording of their panel on Accountability & Integrity in AI (via zkML) here.
There is a lot of excitement about the intersection of these spaces, but it is still early. Many folks in the crypto space are thinking about actionable use cases for what they are building. As we’ve written, we’re excited to see where crypto principles and tools are relevant in other verticals, and AI offers interesting avenues of exploration. But we have to be thoughtful as we proceed; especially when it comes to topics like alignment.
You can find most of the recordings here: https://zuzalu.streameth.org/ by selecting the AI x Crypto track. Note that some were not recorded by request of the speakers.
Special thanks to the different folks who came together to make this happen: Stuti, Vitalik, Janine, Vincent, Yesh, Marine, and many others. Thank you to folks who hosted side events around the summit, like the Flashbots crew and Christian. And thank you to Grimes and Serene for graciously offering their talent and performing during the event.
With love, til next time ❤️
If these topics are interesting to you, please reach out or follow us on Twitter.