Until recently, the contemporary AI industry had been built around a centralized distribution paradigm where machine learning solutions are delivered as a part of cloud-based APIs and software packages deployed on remote servers of AI providers. Now, we are moving toward the next frontier -- decentralized AI that can run and train on local devices or make decisions in decentralized networks like blockchain.
The transition to decentralized AI is enabled by new technologies, such as Google's Federated Learning, that allow for crowd-training of ML algorithms, device-centric AI that runs and trains ML models on mobile devices and the use of AI in DAOs (decentralized autonomous organizations) on blockchain networks. As a venture studio partner specialized in artificial intelligence, entrepreneurs frequently ask me about the future of the industry and what will truly disrupt this space. In this article, I will discuss how decentralized AI works, what potential it has and, more importantly, what benefits can business owners and users extract from it.
Artificial Intelligence And Decentralized Organizations
One of the most exciting innovations of recent years is DAOs that run on Ethereum blockchain. In a nutshell, a DAO is a computer algorithm that implements token ownership rights, contractual obligations and business logic rules (e.g., when to sell, what to sell). When all these things are put together, we get an algorithmic company run via smart contracts that distributes value among its virtual shareholders. Such design is effective in the decentralized distribution of royalties, stock trading, crowd-funding, disbursements of micro-payments, subscriptions payments, prediction markets and more.
AI DAOs emerge when we entrust some or all decision-making responsibilities to AI agents on the blockchain. AI in DAOs can be implemented in several ways. If you are a holder of ownership rights in some DAOs, you can cede your decision making (e.g., yes/no votes) to an AI agent (another smart contract) that will make all decisions for you. Or, in a more radical scenario, we can put AI at the center of the DAO, making it a de facto manager responsible for all organizational and business decisions. For example, imagine an AI DAO for marketing where the AI manager selects the best companies or users to place your ads with. After each marketing cycle, the AI would assess the ROI and adjust its marketing policies accordingly.
In essence, AI DAOs take us to a qualitatively new economic reality. It is a reality where AI software becomes a kind of business administrator that oversees businesses and learns from and competes with other AI managers in the decentralized network. Powered by data flowing from thousands of users, and having access to resources and the ability to amass them, decentralized AIs can become a source of huge economic value for its owners. For example, using generative models (GANs), we can create AI DAOs that trade in their own art, logos, sketches, images or video clips and distribute profits as cryptocurrency tokens to their shareholders.
Furthermore, we can imagine an AI DAO becoming the only shareholder of the accumulated capital. We may see this approach in Terra0, a project involving an augmented self-owned forest proposed by Paul Seidler and Paul Kolling from the University of Arts, Berlin. In the project, forest land ownership is structured as an AI DAO with smart contracts on the Ethereum blockchain. Then, using drones and satellites, the AI DAO can evaluate the wood stock and decide how much and when to sell in the market. Once the project is up and running, the AI DAO can pay out debts to its initial owners and eventually turn a forest into the autonomous, self-owned entity that controls its own resources. Taking this idea further, we can imagine self-owned AV (autonomous vehicles) and robots becoming a normal part of our future economy.
Decentralization With Google's Federated Learning And Device-Centric AI
Centralized AI solutions provided as APIs and cloud-based services are great, but they have certain bottlenecks. Since users access AI features via the network and because ML algorithms involve heavy computations, high latency is often an issue. Also, if you train AI models in a centralized way, it may take more time to improve them. In contrast, decentralized AI can function locally on users' devices, have access to more user data and have no dependence on a network connection, which means less power consumption and minimal latency. Recent advances in decentralized AI have been made thanks to on-device optimization of AI/ML for smartphones and production of dedicated chips for mobile AI and for desktops (e.g., Google's TPU).
Decentralized AI gained powerful momentum in April 2017 after Google announced its new Federated Learning concept. This innovation signals a transition to fully decentralized learning and device-centric AI where machine learning models are trained directly on smartphones of users. Keeping the privacy of user data intact, Google can now outsource AI training to Android users, enabling on-device improvement of shared models. Federated Learning will solve the problem of high-latency and low-throughput connections where users have to connect to remote servers to use ML software. According to Google's Brendan McMahan and Daniel Ramage, "Federated Learning allows for smarter models, lower latency and less power consumption, all while ensuring privacy."
The move toward device-centric AI can also be seen in the release of Google's TensorFlow Lite, a mobile version of a machine learning library fined-tuned to the computational and power constraints of smartphones. In June 2017, Apple followed Google's lead by releasing its Core ML library for iOS devices. The library ships with the optimized general-purpose ML models and tools to convert third-party models into the iOS format. Making models available locally without a network connection will make it easier to develop mobile applications with AI functionality. According to Dave Burke, Google's vice president of engineering for Android, these innovations "will help power the next generation of on-device speech processing, visual search, augmented reality and more."
In the long run, a combination of AI DAOs, device-centric AI and decentralized learning will make AI more democratic and widespread than ever before.