The Rise of Artificial Decentralized Intelligence (ADI)
The current improvement in the speed of Decentralized Peer-to-Peer technologies fueled by the intermittent Bull Market of Cryptocurrencies and their respective Blockchains is generating the productive soil that can give rise to Artificial Decentralized Intelligences (ADIs). The same as with the concerns with Artificial General Intelligence (AGIs), this occurrence or technological singularity most probably will occur and is only a matter of speed of innovation and the specific point in time of deployment.
The thing with ADIs is that, like Blockchains, one single government will have a very difficult task of deactivating this computer system because the nodes will be distributed among many countries and with economic incentives built into the system. The deployed nodes participating in the ADI network you will be rewarded with cryptocurrencies. Then the same as current, or more, applications of AIs will be available on ADIs, but they are not going to be controlled by a single human entity but by a community of validators guided by economic incentives.
Layer 2 blockchains are designed to be massively scalable and are a natural target for the deployment of machine learning algorithms, but probably a more native approach including high-speed calculations is needed. Layer 2 blockchains, suck as Optimism, Arbitrum, and Starkware, have specific programming languages (the most common is Solidity, Cairo is up and coming for StarkWare) that are not suitable for the high-performance computation of AIs. Zero Knowledge (ZK) is the cryptographic generation of short proof that you have some data or a computation has been done without revealing all the details of the data or the computation. ZK proofs to be useful must also be verifiable in a short time. Future high-speed improvements in Zero Knowledge technologies (StarkWare used this ZK technology) will allow high-performance computation into the blockchains. The main issue with blockchain is that you need any computation of transaction to be quickly validated by other nodes, and ZK allows validations that are much faster than the computation per se.
We can think about which machine learning systems are the better candidates to be first migrated into decentralized systems, and this includes:
Recommender Systems: as users consume different items, this is registered, and this is evaluated to suggest future items to consume (technically, you estimate a distance to other items). This type of technology is a good candidate to apply sharding and splitting recommender algorithm data into many nodes. You do not need all user preferences, the past items consumed, to be stored on a single computer. This thesis describes the academic state-of-the-art 2015 of these new systems. A current approach (from 2022), add more in that direction on this paper.
Clustering/Unstructured Classification: given that clustering is the problem of classifying a dataset into spontaneous new categories seems easier to decentralize than structured classification (fixed number of categories). If you imagine the categories as geographic or spatial regions, you see that there is no need for all data points to be stored in a single computer. Here in this paper, we find the authors developed a generalized clustering algorithm, a survey of academic algorithms was also found, and a decentralized clustering algorithm that was applied to brain imagery was recently published.
Now, in this discussion, the missing tool in AI or Machine Learning is the Structured Classifier. Based on a fixed number of categories, the algorithm has to guess to which category a piece of data belongs (for example, is this article written in English or Spanish?). Is closely related to Reinforcement Learning, which is like a closed loop of classifiers to generate actions for robots or games. Deep Learning is a combination of many layers of Structured Classifiers (hence, deep layering) to get a much more sophisticated automated learning experience. The issue with this type of AI tool this that you need a global view of all the training datasets because the output is using a synthetic summarization of the data in the formal of trained weights or variables. You need the training weights to generate the output, the category, the robot action, etcetera.
Matrix multiplication is doing a lot of numerical multiplications and additions. Massive matrix multiplication is the main operation involved in Structured Classifiers, Deep Learning and Reinforcement Learning. As we mentioned before, the verification of these operations (to avoid cheating) is the main challenge that ADIs will face. We envision these three scenarios for Artificial Decentralized Intelligences (ADIs):
Native High-performance Blockchains or Sidechains: When Bitcoin was deemed useless because of the amount of mindless computing that was “wasted” into validating 5 transactions per second, many visionaries proposed that the mining of blocks involves more useful computations. That is a Holy Grail of blockchain puzzles that will help humanity was suggested. This alternative is in that line of work, to participate in the validating of transactions in the blockchain network, you will have to do matrix multiplications and complex machine learning operations that going to be validated by other nodes and, eventually, accepted as part of the mining of the cryptocurrency. This approach is limited to specific operations or static deep learning architectures. Filecoin and other storage blockchains can be seen in this category by just storing the data but without not many or no transformations. In this publication we found, optimization problems have been proposed, but that is not quite a matrix multiplication. WekaCoin solution proposed an array of diverse machine learning algorithms to be involved in the consensus to make the mining more intelligent.
Faster Layer2 Blockchains: Leveraging the existing high-performance and inexpensive Layer 2 blockchains, most of them based on the Ethereum network protocol, is a natural approach to implementing an Artificial Decentralized Intelligence. Using Solidity as a programing language might not be the fastest, but the technology has all the ingredients to build the Decentralized AI Lego, ie. building blocks of reusable machine learning code that is open and free to use (if you pay the network fees). The main limitation of this approach is that blockchains usually have a limited computation power that can be included in a single block. Then, if you fork Layer 2s like Arbitrum, Optimism, or Starkware you have to be ready to increase the maximum block size by a lot and also be ready to put a minimum performance threshold to validators in the network. You do not want commodity validators, you want high-performance validators to process your super computation-intensive blocks.
Specialized Zero Knowledge Platforms for AI: This alternative is similar to the StarkWare Layer 2 approach mentioned before but also involves the specific development of ZK smart contracts (for example using Cairo programming language) for matrix multiplication and deep learning. This can be done in the smart contract layer, such as in StarkWare, or in a lower consensus layer (see zkEVM for a Turing-complete ZK computer). The goal is to have a massive heavy computation that is easily verified by other nodes in the network. Also, including flexible smart contract composition of operations allow for the interoperability of different algorithms.
Will ADIs become the AI robot dictators that are going to rule our lives, like Elon Musk suggested, or are going to be obedient and productive tools of an abundant and less materialistic future?
