No, we’re not a decentralized AI marketplace
Combining two of the hottest trends in tech right now, AI and blockchain, would seem a recipe for success. Indeed, many blockchain projects today marry the two using a decentralized marketplace model, matching datasets to AI developers or AI systems to AI buyers.
Building a marketplace in established markets is tough enough, however, let alone trying to build one in a nascent market such as AI where the use cases and value propositions are not yet clear or deep. Savvy investors have more than once asked us hard questions about the viability of an AI marketplace model.
The Language Network (LangNet) is fundamentally different, we tell them, because the use case and problem to be solved are both very clear: we want voice apps, and we need computers to understand our voice. This requires a layer of common infrastructure more similar to AWS or Filecoin than a marketplace such as Ebay or Airbnb.
Marketplace vs. Resource Network
A marketplace is place where buyers and sellers match up to exchange value. In a marketplace, generally speaking, everything is unique: each buyer has unique needs, each seller has a unique product, and each transaction is self-contained and unique. Airbnb is a great example of a marketplace that matches the unique needs of each traveler with unique accommodations.
A resource network, as we define it, is a market in which the product is a commodity, not unique. The best example of a resource network is Filecoin, a “decentralized storage network.” Storage needs are mostly the same and storage suppliers basically interchangeable. When you have such a liquid, continuous supply side, you are no longer matching buyers to sellers but rather providing a pool of resources that buyers can continually access.
LangNet often gets mistaken for being an AI marketplace because that is how the market for language AI training data operates today. When a developer wants to build a travel app for Japanese, he must acquire a language dataset for travel in Japanese from a data broker or travel company in a one-off transaction. This is a basic “buyer matched with seller” marketplace model.
One of the fundamental insights underpinning LangNet is that, by amassing a large amount of data and using blockchain, we can transform this language data marketplace into a language resource network. We use token incentives to crowdsource a huge, universal set of voice data across 50 different languages and myriad use cases, enabling any developer to quickly bootstrap the data they need to build their app or service.
We can create a resource network instead of an AI marketplace because language AI, unlike many other forms of AI, has clearly defined use cases and common requirements: we know how many languages there are in the world, and we know what use cases we should build datasets for (i.e. travel, banking, healthcare, education, and so on). Well-defined requirements mean that datasets need not be unique to each transaction. They can be reused over and over again, as long as the datasets are broad and deep and aggregated at a system level. Much like storage, language data then becomes a commodity.
Therefore, our approach is to collectivize these resources and create a “language resource network” that buyers can continually access. We spent much time thinking about the design of our system to accomplish this goal. It comes down to three mechanics:
- We use a mining schedule to build out the data supply side. Mining effectively means that the ecosystem is “pre-paying” for language data and infrastructure that anyone can use, similar to the way our governments “pre-paid” for roads and bridges.
- Resource providers such as data validators and data hosts get paid through token inflation, which means the ecosystem as a whole pays for these services much as our taxes collectively pay for society-level infrastructure and services.
- Resource consumers stake tokens for continuous access to the whole platform’s resources rather than pay for one-off transactions.
There is no matching of buyers and sellers, no unique transactions, just a continous, liquid supply of common resources. More details are available in our whitepaper.
The other way we think about LangNet is as voice dApp platform. Building your own datasets and models to develop a new voice app leads is akin to building your own servers for every new website, or a new blockchain protocol for every new dApp. There is much redundant, wasted effort.
As AWS resolved this for web apps and Ethereum is addressing for dApps, one of the goals of LangNet is to commoditize language AI so developers can focus on building new voice dApps instead.