Coinbase Machine Learning and Blockchain 

Updated: May 19, 2024

My experience

This month, I attended the live virtual Machine Learning and Blockchain Research Summit. What immediately struck me was the emphasis on academic “research.” And it indeed showed with the presenters’ credentials and types of their talks. The full list of speakers can be found on the website https://www.coinbase.com/ml-bc-summit, to name a few: Salman Avestimehr, professor of electrical & computer engineering (ECE) and computer science (CS) at University of Southern California, Yupeng Zhang, assistant professor of ECE  and CS at the University of Illinois Urbana-Champaign, and Bhaskar Krishnamachari, professor of ECE  and CS at the University of Southern California. Of course, there were prominent speakers from the industry, including Vitalik Buterin. 

Indeed, the discussion with Vitalik was quite interesting as one can sense a deep thought process in formulating a deep answer for not a superficial question. I enjoyed it as much as the panel with Sam Green on it. Sam’s analogies, like the blockchain to the brain that is open to everyone to see, were so illustrative that I would borrow them for my undergraduate class. Of course, the expert knowledge of all speakers was well-appreciated, especially in learning about the peculiar technical challenges that oftentimes slip researchers’ attention when modeling generalized incentive mechanisms or theoretical models.  

Key Takeaways

As I listened to the presentations and speakers, several ideas were carved more distinctly in my mind:

• Big picture. The point of the summit was to hypothesize the ways in which blockchain can benefit from AI and vice versa, how AI can benefit from blockchain.

• Blockchain-AI symbiosis. The main mentioned pros for the AI from properties of public blockchain such as transparency, immutability, verifiability, and pseudonymity are:

  • Data verification. Contemporary AI systems lack the data quality control. Going forward the problem of the legitimacy and truthfulness of data will exacerbate as more generated content will appear in the public media that are used to train the next gen AI (kind of uroboros). Therefore, blockchains can aid by allowing to store one kind of data (e.g., truthful, real) and providing proofs of the data authenticity upon request.  

  • Data ownership. Related to the previous point, user data can be tokenized and linked to a particular owner ID. This way any piece of data can be traced back to the owner for rewards, verification, confirmation, etc. 

  • Computation power. Besides the data, AI systems extensively use computing power. The majority of the modern state-of-the-art AI models are private in the sense that the computing capacity used for training is centralized under one organization, typically private. Therefore, the governance over the model setup and training is centralized. Another angle of the problem is the high barrier to entry into the AI market due to market monopolization. Big centralized players with a lot of market and financial power monopolize the CPUs and block competitors from entering the market. The decentralized computing model used in the proof-of-work blockchains can be repurposed to provide computer power on demand to others. Pooling of limited individual processing capacities may extend the limits for alternative players. This approach is also called federated learning and is described in detail here. 

  • Cryptography. Undoubtedly, many speakers admitted that the main enabling feature of blockchain for AI is cryptography. For example, data integrity across the ETL process could be reinforced by cryptography, thus maintaining accuracy and consistency throughout the whole processing chain. Nevertheless, cryptography too has its limitations in remedying all issues with AI models.

  • Limitations. Where blockchain is not suitable in its current state for AI is the memory part. Data storage on blockchain is not sufficient for (quality) ML.

• AI-Blockchain symbiosis. The consensus among speakers was that the blockchain needs AI, but AI does not necessarily need blockchain. The current blockchain infrastructure is lacking UX/UI for which the Gen AI could be quite a propeller of adoption.

  • User experience. One of the current obstacles to the mass adoption of blockchain technology is the lack of user experience. AI may provide a convenient conversational interface for blockchain applications. For example, one solution presented was a conversational AI that interprets the human query into SQL code that extracts data from the blockchain transaction database. Nevertheless, very few interesting use cases of AI as an interface were presented at the summit. As Vitalik wittingly pointed out, the app layer of businesses, such as Coinbase, should probably be the ones spearheading this innovation, contrary to what developers at Ethereum Foundation are doing, for example.

• Incentives. The final point I wanted to mention is the references to the incentive structures throughout the sessions. If the decentralization is meant to involve a network of actors, the incentives for the provision of their resources and coordination with other actors should be at the center of attention for builders. The key objective – is to have a reliable computing infrastructure, which is sustained in the long run only if the parties are incentivized. We will continue seeing centralized platforms, like Azure or AWS, play a big role in the blockchain and AI future without incentives for individual computing power and data providers. In addition, the speakers stayed agnostic about the open or closed-source future of the blockchain and AI. 

If you were to watch one segment from this event, I would recommend this panel: https://vimeo.com/event/4171179#chapter=15081729

The whole recording of the event can be found here: https://vimeo.com/event/4171179