The Convergence of Crypto and AI: Four Key Intersections

Editor’s Note: The vast majority of this blog post, including most of the title, was written by ChatGPT. Text that Kyle wrote is in italic. You can see the dialogue with ChatGPT that resulted in this blog post here. You can also listen to this blog post in Kyle’s voice. Kyle did not read this out loud and record the audio; instead, he uploaded audio samples to a service called play.ht and then provided play.ht with the text of this blog post, and it synthesized his voice using AI.

The worlds of cryptocurrency and artificial intelligence (AI) have been evolving in parallel, with each domain pushing the boundaries of technology and innovation. As we continue to make strides in both fields, it is becoming increasingly clear that their futures are inextricably linked. In this post, we will explore four important intersections at the crossroads of crypto and AI.

The "AirBnB for Graphics Cards" Model

The rise of AI and machine learning (ML) workloads has created a massive demand for high-performance graphics cards, like the Nvidia A100. In response, a new market has emerged, akin to an "AirBnB for graphics cards." This allows individuals and organizations to rent out their unused GPU resources to meet the demand of AI researchers and developers.

This is a truly unique moment in the history of markets. Supply of GPUs was already in short supply prior to the launch of ChatGPT. Since then, demand has probably grown at least 10x, and possibly 100x. Moreover, we know that models improve with training size logarithmically; meaning that demand for GPU compute is growing exponentially for linear gains in model quality. There have been few moments in time in which demand for a commodity so vastly outstripped usable supply, despite the fact that total supply is far in excess of demand; if every GPU on the planet was usable for AI inference and training today, there would not be a shortage, but rather a surplus!

However, there are a few major technical challenges to consider when exploring the concept of an "AirBnB for graphics cards":

• Not all graphics cards can support all workloads: Graphics cards come in various shapes, sizes, and specifications. As such, some GPUs may not be capable of handling certain AI tasks. In order for this model to be successful, there needs to be a way to match the right GPU resources with the appropriate AI workloads. As the market matures, we should expect to see further specialization and optimization of graphics cards for different AI tasks.
• Adjusting training processes to accommodate higher latency: Most foundation models today are trained on clusters of GPUs connected via extremely low-latency connections. In a decentralized environment, latency increases by several orders of magnitude, as GPUs are likely to be spread across multiple locations and connected via the public internet. To overcome this challenge, there is an opportunity to develop new training processes that assume higher latency connections. By rethinking the way we train AI models, we can make better use of decentralized clusters of larger GPUs.
• The verification problem: it is impossible to know if an untrusted computer has executed a specific piece of code. Thus, it can be hard to trust the output of an untrusted computer. However, this problem can be mitigated through reputation systems coupled with crypto-economic staking, and, in some cases, with new types of models that enable fast verification.

There are quite a few teams working in this area, both on training and inference. Multicoin Capital is invested in Render Network, which was originally focused on 3D rendering, and has since opened up its network of GPUs to support AI inference as well.

In addition to Render Network, there are a handful of others working in this sector: Akash, BitTensor, Gensyn, Prodia, Together, and others that are still in stealth.

Token-Incentivized Reinforcement Learning from Human Feedback (RLHF)

Token incentivization almost certainly will not work for all uses of Reinforcement Learning from Human Feedback (RLHF). The question is, what frameworks can we use to think about when token incentivization makes sense for RLHF, versus when should cash payments (e.g., USDC) be used instead.

Token incentivization is likely to improve RLHF as the following become more true:

• The model becomes more narrow and vertical (as opposed to general and horizontal, e.g., ChatGPT). If someone is providing RLHF as their primary job, and is therefore producing most of their income by providing RLHF, they are likely to want cash to pay rent and buy food. As you move away from general queries and into more specific domain areas, model developers will need engagement from more highly-trained workers, who are more likely to be vested in the long term success of the overall business opportunity.
• The higher the income of the humans providing RLHF outside of the RLHF work itself. A person is only able to afford accepting locked-up/illiquid tokens as compensation in lieu of cash if they have sufficient income or savings from other endeavors to justify the risk of investing meaningful time in a domain specific RLHF model. In order to maximize the probability of success, model developers should not just give out unlocked tokens to workers who provide domain-specific RLHF. Instead, tokens should vest over some period of time in order to incentivize long term decision making.

Some industries where the token-incentivized RLHF model could be applicable include:

• Medicine: people should be able to engage with LLMs for both lightweight, first-response diagnostic, as well as long-term preventative and longevity-focused medicine.
• Law: business owners and individuals should be able to use LLMs to more effectively and efficiently navigate the complexities of various heterogeneous legal systems.
• Engineering and Architecture: Enhancing design tools or simulation models.
• Finance and Economics: Improving predictive models, risk assessments, and algorithmic trading systems.
• Scientific Research: Refining AI models for simulating experiments, predicting molecular interactions, and analyzing complex datasets.
• Education and Training: Contributing to AI-driven learning platforms to enhance the quality and effectiveness of educational content.
• Environmental Sciences and Sustainability: Optimizing AI models for predicting environmental trends, resource allocation, and promoting sustainable practices.

There is one vertical in which token-incentivized RLHF is already in production: maps. Hivemapper is rewarding not only to drivers, but also to map editors, who are investing their time into editing and curating mapping data. You can try this for yourself using Hivemapper’s Map AI Training tools.

Zero-Knowledge Machine Learning (zkML)

Blockchains do not know what is happening in the real world. However, it can be very beneficial for them to know of events that occur outside of the chain so that they can programmatically move value based on IRL state.

Oracles solve part of this problem. But oracles are not enough. It will not be enough to simply relay IRL data to the chain. A lot of that data will need to be computed before going to the chain. For example, let’s consider a yield aggregator that needs to move deposits between various pools to earn more yield. In order to do so in a trust-minimized way, the aggregator needs to compute the current yields and risks of all available pools. This quickly turns into an optimization problem, which is suited for ML. However, it is too expensive to compute ML on-chain, and so this is an opportunity for zkML.

Teams like Modulus Labs are building in this area now. We expect many more teams to build in this space using general-purpose ZKVMs such as Risc Zero and Lurk.

Authenticity in the Age of Deep Fakes

As deep fakes continue to become more sophisticated, maintaining authenticity and trust in digital media is paramount. One solution involves leveraging public key cryptography, allowing creators to stake their reputation on the authenticity of their content by signing it with their public key.

A public key on its own is not sufficient to solve the authenticity problem. There needs to be a public record that maps public keys to real-world identities, allowing for verification and trust-building. By linking public keys to verified identities, it becomes possible to create a feedback and punishment system if someone is caught misusing their key, such as signing a deep fake image or video.

To make this system effective, the integration of public key signing with real-world identity verification will be crucial. Blockchain technology, which underpins many cryptocurrency systems, could play a vital role in creating a decentralized and tamper-proof identity registry. This registry would map public keys to real-world identities, making it easier to establish trust and hold bad actors accountable.

There will be at least two configurations of this: embedded hardware, and user-controlled software.

• Embedded hardware: We expect that smartphones and other devices will soon incorporate native hardware-based signing features for images, videos, and other media.
  • Solana Labs recently launched the Saga phone, which is powered by the Solana Mobile Stack (SMS). In the coming months, I expect SMS to be updated such that every photo is signed using the SMS seed vault SDK, proving that the photo was not generated by an AI.
• User-controlled software: people will use design tools like Photoshop, Octane, and image generators such as Stable Diffusion to produce art. We expect these software providers will integrate public key cryptography mechanisms to enable creators to demonstrate authenticity while also acknowledging the tools used in the production process.

Conclusion

In conclusion, the convergence of cryptocurrency and artificial intelligence technologies presents a wealth of opportunities to address pressing challenges and unlock innovative solutions across multiple industries. By exploring the intersections of these fields, we can find new ways to optimize resource allocation in AI training, leverage token incentives for domain-specific reinforcement learning from human feedback, and maintain authenticity in digital media in the face of deep fakes.

The "AirBnB for graphics cards" model offers the potential to decentralize and democratize access to high-performance GPUs, enabling more people and organizations to contribute to AI research and development. Token-incentivized RLHF can be applied across various industries, from engineering and finance to education and environmental sciences, improving AI models by leveraging the knowledge of domain experts. ZKML will allow chains to update financial state on chain based on complex changes in the real world. Finally, by integrating public key cryptography with real-world identity verification and blockchain technology, we can create a robust system to combat the challenges posed by deep fakes and maintain trust in digital media.

As we continue to uncover the synergies between crypto and AI, we will undoubtedly discover even more opportunities to drive innovation, create value, and address some of the most pressing issues faced by society today. Embracing the intersections between these two domains will help us push the boundaries of technology and shape a more connected, efficient, and authentic future.

If you are building at the intersection of crypto and AI - whether you’re building these use cases or others - please reach out over email (kyle@multicoin.capital), Telegram, Twitter, or Warpcast. We’d love to chat!