The Convergence of Crypto and AI: Four Key Intersections

Author: Kyle Samani (Multicoin Capital Partner) and ChatGPT; Translation: Jinse Finance cryptonaitive and ChatGPT

*Note: The vast majority of this article, including most titles, was written by ChatGPT. Text written by the author is in italics. You can see the author's conversation with ChatGPT at here. *

The worlds of Crypto and AI have been developing in parallel, with each field pushing the boundaries of technology and innovation. As we continue to make progress in both fields, it becomes increasingly clear that their futures are closely intertwined. In this post, we will explore four important intersections at the Crypto and AI crossroads.

"AirBnB for graphics cards" model

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

*This is a truly unique moment in the history of the market. Before the launch of ChatGPT, the supply of GPUs was already in short supply. Since then, demand has probably grown at least 10-fold, possibly 100-fold. Furthermore, we know that models grow logarithmically with training size; this means that the need for GPU computing increases exponentially to improve model quality. While total supply far exceeds demand, moments when demand for a commodity so massively outstrips available supply are rare; if every GPU on the planet today could be used for AI inference and training, instead of a shortage, there would be excess! *

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

• Not all graphics cards support all workloads: Graphics cards come in all shapes, sizes and specifications. Therefore, certain GPUs may not be able to handle certain AI tasks. In order for this model to be successful, there needs to be a way to match the correct GPU resources with the appropriate AI workload. As the market matures, we should expect to see further specialization and optimization of graphics cards for different AI tasks.

• Tuning the training process for higher latency: Most base models today are trained on clusters of GPUs connected via extremely low latency. In a decentralized environment, latency increases by orders of magnitude, as GPUs may be distributed across multiple locations and connected via the public internet. To overcome this challenge, there are opportunities to develop new training procedures with higher latency connections. By rethinking the way we train AI models, we can take better advantage of decentralized clusters of larger GPUs.

• *Verification Problem: There is no way of knowing whether a particular piece of code was executed by an untrusted computer. Therefore, it is difficult to trust the output of an untrusted computer. However, this problem can be mitigated by reputation systems combined with cryptoeconomic staking and, in some cases, by novel models that support fast verification. *

• There are quite a few teams working in this area, both training and inference. MulticoinCapital invested in Render Network, which initially focused on 3D rendering and has opened up its GPU network to also support AI inference. *

*Besides RenderNetwork, there are a few other companies working in this space: Akash, BitTensor, Gensyn, Prodia, Together, and others still in development. *

Token incentive RLHF (Reinforcement Learning from Human Feedback, reinforcement learning based on human feedback)

Token incentives will almost certainly not work for all use cases of reinforcement learning from human feedback (RLHF). The question is, what framework can we use to consider when token incentives make sense for RLHF and when cash payments (e.g. USDC) should be used.

Token incentives may improve RLHF as the following becomes more true:

• The model has become more narrow and vertical (as opposed to generic and horizontal, e.g. ChatGPT). *If someone will provide RLHF as their main job and therefore generate most of their income from providing RLHF, they may need cash to pay rent and buy food. As you move from general queries to more specific domains, model developers will require the involvement of more trained staff who are more likely to have long-term success in the overall business opportunity. *
• ** Higher income for those who provide RLHF outside of the RLHF job itself. ***One can only accept locked/illiquid tokens as compensation if one has sufficient income or savings from other endeavors to justify the risk of investing meaningful time in a domain-specific RLHF model Not cash. To maximize the probability of success, model developers should not only issue unlocked tokens to workers who provide domain-specific RLHF. Instead, tokens should be vested over a period of time to incentivize long-term decision-making. *

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

*Medicine: *One should be able to practice lightweight, first responder diagnostics, as well as long-term preventive and longevity medicine with an LL.M. *

• Legal: * Business owners and individuals should be able to use LLM to more effectively navigate the complexities of various heterogeneous legal systems. *
• Engineering and Architecture: Enhance design tools or simulation models.
• Finance and Economics: Improve forecasting models, risk assessment and algorithmic trading systems.
• Scientific Research: Improve AI models for simulating experiments, predicting molecular interactions, and analyzing complex datasets.
• Education and Training: Contribute to AI-powered learning platforms to improve the quality and effectiveness of educational content.
• Environmental Science and Sustainability: Optimizing AI models to predict environmental trends, resource allocation, and promote sustainable practices.

There is one vertical where token-incentivized RLHF is already in production: Maps. Hivemapper is not only good for drivers, but also for map editors who invest their time editing and organizing map data. You can try the map AI training tool yourself using Hivemapper.

Zero Knowledge Machine Learning (zkML)

*Blockchain has no knowledge of what is happening in the real world. However, it would be very beneficial for them to understand what is happening off-chain so they can transfer value programmatically based on real-world state. *

• Oracles solve part of this problem. But oracles are not enough. Simply relaying real-world data onto the chain is not enough. Before entering the chain, a lot of data needs to be calculated. For example, let's consider a yield aggregator that needs to transfer deposits between different pools to earn more yield. To do this in a trust-minimized manner, the aggregator needs to calculate the current payoff and risk for all available pools. This quickly becomes an optimization problem suitable for ML. However, computing ML on-chain is too expensive, so this is an opportunity for zkML. *

*Teams like Modulus Labs are building in this space right now. We hope more teams are building in this space using general-purpose ZKVMs, such as Risc Zero and Lurk. *

Authenticity in the age of deepfakes

As deepfakes become more sophisticated, maintaining authenticity and trust in digital media is critical. One solution involves utilizing public-key cryptography, allowing creators to guarantee the authenticity of their content by signing it with a public key.

The public key alone is not enough to solve the authenticity problem. There needs to be a public record that maps public keys to real-world identities for verification and trust. By associating public keys with verified identities, a system of feedback and penalties can be created if someone is caught abusing their keys, such as signing deepfake images or videos.

For this system to be effective, the integration of public key signatures with real-world authentication will be critical. Blockchain technology, which underpins many cryptocurrency systems, can play an important role in creating decentralized and tamper-proof identity registries. The registry maps public keys to real-world identities, making it easier to build trust and hold bad actors accountable.

• There will be at least two configurations: embedded hardware and user-controlled software. *

• Embedded Hardware: We expect smartphones and other devices to soon integrate hardware-based native image, video and other media signing capabilities.

*SolanaLabs recently launched Saga Phone, which is powered by Solana Mobile Stack (Solana Mobile Stack, SMS). In the next few months, I hope that SMS will be updated so that each photo is signed using the SMS Seed Bank SDK, proving that the photo was not generated by AI. *

• USER-CONTROLLED SOFTWARE: People will use design tools like Photoshop, Octane, and image generators like StableDiffusion to make artwork. We expect these software providers will integrate public-key cryptography, enabling creators to prove authenticity while also acknowledging the tools used in the production process.

in conclusion

In conclusion, the convergence of cryptocurrency and AI technologies offers ample opportunities to address pressing challenges and unlock innovative solutions across multiple industries. By exploring the intersection 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 keep digital media authentic in the face of deepfakes sex.

The "AirBnB of 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 industries ranging from engineering and finance to education and environmental science, improving AI models by leveraging the knowledge of domain experts. ZKML will allow the blockchain to update the financial state on-chain based on complex changes in the real world. Finally, by combining public-key cryptography with real-world authentication and blockchain technology, we can create a robust system to meet the challenges posed by deepfakes and maintain trust in digital media.

As we continue to discover the synergy between encryption and artificial intelligence, we will undoubtedly discover more opportunities to drive innovation, create value, and solve some of the most pressing problems facing society today. Embracing the intersection between these two fields will help us push the boundaries of technology and shape a more connected, efficient and authentic future.