Original Source Here
What are buyers buying?
In this past article, we gave a pretty succinct explanation of the asset that sellers are selling (alternatively, what buyers are buying). However, it’s probably a good idea to discuss more high-level where Volant AI models fit in the spectrum of what could be sold in an AI model marketplace.
In terms of abstraction, there’s (probably) four levels to assets that could be sold:
- E2E (End-to-End) solutions: A dashboard tied with the model. Metrics, KPIs, usage, full functionality/integration details with your respective framework.
- Widgets: A plugin/piece of code put into a site
- Inference APIs: A wrapper (of sorts) around the predict functionality of a model. Our models at Volant AI.
- Model Files: File(s) with architecture/weights of a model.
Even though we are an AI model marketplace, the “models” we are selling are docker images that are super-secure Inference APIs. We do not want to sell E2E solutions (or even Widgets) since we are a largely generic AI model marketplace that wants to sell ready-to-use assets. An E2E solution may not be too difficult to integrate, but it is likely far more costly than the other options given their added functionality. Make no mistake, for highly specialized problems, such E2E solutions are necessary.
Widgets are an interesting solution, but this may lead to lock-in with certain platforms. In other words, the model assets that will be uploaded by sellers will need to be turned into specific code for slack, salesforce, chrome, etc. which is not necessarily generalizable. Their use will also be governed by the problem itself and its relevance to specific model/domain which is hard to apply generically. This certainly may be something we need to develop in the future, however.
While Model files (for Keras, the .h5 binary file) can be sold, there’s also some level of framework familiarity needed for usage, especially for those not familiar with deep learning. At the moment, it seems more prudent to not be so low-level and rather allow people to use the model-as-an-API that they can build upon. Clearly, those using our models need to be familiar with how to use APIs, but staking ourselves at this abstraction level lets us potentially move up if need be (perhaps selling a model-as-a-GUI/UI to less technically-inclined folks).
A note on E2E solutions
We’re actually open to helping people sell E2E solutions — in our nomenclature, external 3rd party models — on our marketplace and have even set up the upload model process to make this as easy as possible. The downside of having such assets on our marketplace will be that we aren’t as responsible for the quality control and will not be able to integrate the Model-Selection-as-a-Service component that differentiates us from our competitors. In those situations, we are functioning as hands-off brokers.
Trending AI/ML Article Identified & Digested via Granola by Ramsey Elbasheer; a Machine-Driven RSS Bot