Why does AI need decentralization?

As we proudly pronounce that we live in an “AI (Artificial Intelligence) — first age” and rely increasingly on AI infused products and services to provide us recommendations, feed us information, or influence our opinions at a personal and public level, should we ever pause and give it a minute to think about are the AI models that serve us everyday reliable and trustworthy enough for us to allowing it to play a critical role in our daily lives?

Digging a little bit deeper, we realize most AI models and algorithms are owned and controlled by centralized organizations, or tech giants. It may be scary to find out how much they know about you, or how little they actually know about you to tag labels on you in their gigantic database.

It sounds like a terrible idea to let some centralized organizations to collect and monetize our data, having the full control and processing these data in some black boxes. Our data will become increasingly valuable as things are slowly becoming more automated, not only that we are just giving them for free, but we also have no clue if these organizations have a complete picture of us in order to produce unbiased AI models that serve in all the products we use.

Biased and siloed data sets

Speaking of collecting data, as much as large companies try to extract valuable data from multiple sources, such as transaction logs, social media interactions, search behaviors, clicks from products, or sensor and machine logs, most of these data sets are not only messy, but also siloed from each other, and thus not ideal for training advanced AI models.

Expensive and wasted computing resources

Technically speaking, we question how effective it is for one centralized organization to own all AI models. Take an example of self-driving cars, if there are three cars that stopped almost the same time at one intersection, and they are trying to figure out who should go first. In today’s centralized world, these cars would have to communicate to some centralized server to give them recommendations on how to proceed, then it would not only be time consuming, but also a waste of computational resources. Ideally, these cars should have the ability to talk to each other to quickly figure out who has the right-of-way.

Higher threshold makes it hard for smaller players to enter the market and limits the innovation of AI

We’ve conducted researches with companies who would like to incorporate AI into their products and services. Yet the outlook is quite iffy with the increasing higher threshold for smaller players to enter the market or benefit from existing AI models. Just to name a few macro hurdles that these companies are facing:

• Access to AI models is very limited
• AI models shared from tech giants are not customized or optimized for individual needs
• Lack of trust around data exfiltration and system authentication
• Too expensive and time consuming to adopt AI models released by tech giants

After learning about various challenges for relatively smaller companies to adopt open sourced AI models from tech giants, we saw a huge need for the industry to have an open and trusted platform, where all participants can collaboratively work together to advance AI models, get fair share of rewards based on contribution, and at the same time able to mitigate privacy and security issues. The CEO of TuringNet, Dr. Kai Wen, who was a former Senior Engineer from Google, also a PhD from Stanford University, started tapping into how Blockchain technology can be applicable here to serve as an infrastructure for such platform that performs distributed computing for AI model training at scale in an open and trusted environment.

The idea behind TuringNet is aiming to solve industry-wise problems such as opaque data refinery or concerns around data exfiltration, siloed and fragmented data sets, and repetitive model training that wastes massive computational resources. With this decentralized and collaborative approach, participants on the platform are incentivized to contribute either computation power or valuable datasets to further advance AI models, developers can also get rewarded for making their AI algorithms accessible for commercial usage. We believe TuringNet will empower a variety of optimized AI models that can be easily adopted by hundreds of millions of use cases in the long term.