Decentralized AI on Blockchain
AI needs to be decentralized and personalized, and to do that you need blockchain.
What’s wrong with AI without blockchain? Let’s look at a well known example of AI.
Google’s centralized AI is powerful but also dumb at the same time. Powerful because it has pushed machine learning based on generic mass data to an amazing level, but dumb because it has virtually no learning capability at client/personal level.
The Google Assistant, for example, is really NOT YOUR assistant, as it does not know your personal spheres and preferences at all, and there is no way for you as the user to even manually teach it about you, much less to hope it automatically learn from you.
Case in point: Every day, I use Google Assistant’s “read it to me” and get amazed by its better-than-average-human reading capability, but at the same time peeved by its persistent inability to reliably recognize my simple and clear voice instruction “read it to me!”
I’d have a webpage open, and say “hey Google, read it to me,” but instead of starting to read the webpage as it is supposed to, Google Assistant would often do completely irrelevant things such as performing a Google search and announcing to me, “here are the things to read,” or going to a website and declaring “here are the things I found on Reddit,” etc.
It continues to do that regardless of how many times I’ve used the same clear and simple instruction with the same voice in similar circumstances.
It is not that Google Assistant doesn’t work at all. In fact, it works most of the time, but just not reliably nor intelligently. When it does get my instruction and start to read the webpage aloud, wow, what a read, better than an average real person’s natural reading, certainly better than mine. But that is a different point. I’m talking about an AI assistant that is supposed to learn, from me the user, but Google Assistant simply doesn’t.
No, I’m not naïvely expecting speech recognition to be perfect either. I’ve used speech recognition professionally for many years, and I fully understand the limitations of the technology.
The point I’m making here is that Google’s acclaimed AI actually does not learn at all at the user level. If it did even just a little bit, it would have learned that when this particular user (me) opens a webpage and says something even vaguely sound like “read it to me”, he means to ask the Assistant to read that page aloud, rather than to search Google or go to Reddit (clearly a misrecognition of “read it”), or any other random things.
If the user intention isn’t clear after just a few times, how about hundreds of times, and how about repeating it every day?
Imagine an assistant or an agent you hire to work for you, who never actually learns anything about you, but rather stays preoccupied by his own fixed routines, and worse, always pretends being smart for doing that.
So it is a fact: The world’s best AI does not learn anything from me the user. The emperor really does not wear any cloth. Everything is pre-trained as far as the user level is concerned. All the machine learning happens at the server level based on aggregated mass data.
This is because Google Assistant is centralized. It sees me the user as a mere data point fitting a general model.
People complain that Google knows too much about users, but the reality is that Google knows both too much and too little.
AI needs to be decentralized and personalized.
So AI needs to be decentralized and personalized.
And for that there is Fetch.AI, a UK Cambridge-based company started by a team that had deep roots in DeepMind.
Fetch.ai is one of the most important decentralized AI experiments being carried out on blockchain now, but seems to receive less recognition than it deserves.
Fetch.ai isn’t cheap piling up of buzzwords (AI + blockchain). It is a serious attempt to find the truth in balancing centralization and decentralization in AI.
Fetch AI is decentralizing AI using blockchain technology. Blockchain does not directly contribute to AI computation itself, but enables decentralized transactions (an all-encompassing concept) which provide an economic foundation for decentralized AI. With decentralization, AI becomes better, more relevant, more intelligent and more useful.
In addition to the benefit of personalization, decentralization is also the most efficient (perhaps even the only practical) AI architecture to connect data silos and place them under a shared common AI framework.
Personalization alone means Fitch.ai deserves a lot more attention. Even if Fetch.ai may deliver only a part of all that it promises, at least the autonomous economic agents (AEA’s) are going to be your assistants due to personalized machine learning.
I look forward to my robo-agents that actually learn from me and work for me.