This week’s AccelNow Blog highlight comes from Medium.com’s Francesco Corea.
Read the article in its entirety Here.
Corea prepared a ton of research for this article, including examples of companies doing good and bad jobs of changing the landscape of the 21st century.
How AI could change Blockchain
Although extremely powerful, a blockchain has its own limitations as well. Some of them are technology-related while others come from the old-minded culture inherited from the financial services sector, but all of them can be affected by AI in a way or another:
- Energy consumption: mining is an incredibly hard task that requires a ton of energy (and then money) to be completed (O’Dwyer and David Malone, 2014). AI has already proven to be very efficient in optimizing energy consumption, so I believe similar results can be achieved for the blockchain as well. This would probably also result in lower investments in mining hardware;
- Scalability: the blockchain is growing at a steady pace of 1MB every 10 minutes and it already adds up to 85GB. Satoshi (2008) first mentioned “blockchain pruning” (i.e., deleting unnecessary data about fully spent transactions in order to not hold the entire blockchain on a single laptop) as a possible solution but AI can introduce new decentralized learning systems such as federated learning, for example, or new data sharding techniques to make the system more efficient;
- Security: even if the blockchain is almost impossible to hack, its further layers and applications are not so secure (e.g., the DAO, Mt Gox, Bitfinex, etc.). The incredible progress made by machine learning in the last two years makes AI a fantastic ally for the blockchain to guarantee a secure applications deployment, especially given the fixed structure of the system;
- Privacy: the privacy issue of owning personal data raises regulatory and strategic concerns for competitive advantages (Unicredit, 2016). Homomorphic encryption (performing operations directly on encrypted data), the Enigma project (Zyskind et al., 2015) or the Zerocash project(Sasson et al., 2014), are definitely potential solutions, but I see this problem as closely connected to the previous two, i.e., scalability and security, and I think they will go pari passu;
- Efficiency: Deloitte (2016) estimated the total running costs associated with validating and sharing transactions on the blockchain to be as much as $600 million a year. An intelligent system might be eventually able to compute on the fly the likelihood for specific nodes to be the first performing a certain task, giving the possibility to other miners to shut down their efforts for that specific transaction and cut down the total costs. Furthermore, even if some structural constraints are present, a better efficiency and a lower energy consumption may reduce the network latency allowing then faster transactions;
- Hardware: miners (and not necessarily companies but also individuals) poured an incredible amount of money into specialized hardware components. Since energy consumption has always been a key issue, many solutions have been proposed and much more will be introduced in the future. As soon as the system becomes more efficient, some piece of hardware might be converted (sometimes partially) for neural nets use (the mining colossus Bitmain is doing exactly this);
- Lack of talent: this is leap of faith, but in the same way we are trying to automate data science itself (unsuccessfully, to my current knowledge), I don’t see why we couldn’t create virtual agents that can create new ledgers themselves (and even interact on it and maintain it);
- Data gates: in a future where all our data will be available on a blockchain and companies will be able to directly buy them from us, we will need help to grant access, track data usage, and generally make sense of what happens to our personal information at a computer speed. This is a job for (intelligent) machines.
How Blockchain can change AI
In the previous section, we quickly touched upon the effects that AI might eventually have on the blockchain. Now instead, we will make the opposite exercise understanding what impact can the blockchain have on the development of machine learning systems. More in details, blockchain could:
- Help AI explaining itself (and making us believe it): the AI black-box suffers from an explainability problem. Having a clear audit trail can not only improve the trustworthiness of the data as well as of the models but also provide a clear route to trace back the machine decision process;
- Increase AI effectiveness: a secure data sharing means more data (and more training data), and then better models, better actions, better results…and better new data. Network effect is all that matter at the end of the day;
- Lower the market barriers to entry: let’s go step by step. Blockchain technologies can secure your data. So why won’t you store all your data privately and maybe sell it? Well, you probably will. So first of all, blockchain will foster the creation of cleaner and more organized personal data. Second, it will allow the emergence of new marketplaces: a data marketplace (low-hanging fruit); a models marketplace (much more interesting); and finally even an AI marketplace (see what Ben Goertzel is trying to do with SingularityNET). Hence, easy data-sharing and new marketplaces, jointly with blockchain data verification, will provide a more fluid integration that lowers the barrier to entry for smaller players and shrinks the competitive advantage of tech giants. In the effort of lowering the barriers to entry, we are then actually solving two problems, i.e., providing a wider data access and a more efficient data monetization mechanism;
- Increase artificial trust: as soon as part of our tasks will be managed by autonomous virtual agents, having a clear audit trail will help bots to trust each other (and us to trust them). It will also eventually increase every machine-to-machine interaction (Outlier Ventures, 2017) and transaction providing a secure way to share data and coordinate decisions, as well as a robust mechanism to reach a quorum (extremely relevant for swarm robotics and multiple agents scenarios). Rob May expressed a similar concept in one of his last newsletters (that I highly recommend — you should definitely subscribe);
- Reduce catastrophic risks scenario: an AI coded in a DAO with specific smart contracts will be able to only perform those actions, and nothing more (it will have a limited action space then).
In spite of all the benefits that AI will receive from an interaction with blockchain technologies, I do have one big question with no answer whatsoever.
AI was born as in an open-source environment where data was the real moat. With this data democratization (and open-source software) how can we be sure that AI will prosper and will keep being developed? What would be the new moat? My only guess at the moment? Talent…
The Article appeared originally on Medium.com by Francesco Corea.
Read the article in its entirety Here.