The document describes how ML product business does not entirely resemble software product business. ML product development needs different philosophies and approaches. In the traditional software product business, companies develop products once and sell multiple times. It is profitable and software companies have ambitions to build ML products using the same business model.
To be successful in ML business, one should be aware of challenges and friction points building ML products. Being aware of these challenges help define ML products.
Writing code once, packaging a product and selling many times is a compelling business model. ML solutions are built by writing code but the heart of these is customer data. The traditional product model does not work for “AI for enterprise”.
Customizing models for each dataset is not a compelling business model. Writing code once lead to building non-AI products.
Most ML models require a “human in loop” approach. Human in loop limit automation. It limit ability to “build once and sell thousand times”
Depending on type of data, volume of data, velocity of data - we select appropriate architecture. There are opinions or reference architecture for different situations. Software products use these. ML intelligent components depend on the content of data. Plurality of methods and approaches in ML is very high and it is hard to form one opinion. It makes it harder to build ML products.
Software products add code/technical capabilities to achieve differentiation. ML capabilities are based on data. In the ML domain, technical differentiation is hard to achieve. Open source modeling architecture is available to all customers.
Product companies want to build products for most/large number of customers. In the ML domain, the generality of the ML model depends on a variety of data. To build a universal solution, one needs universal data. To serve a larger market (customer base), one needs data relative to breadth of market.
Most products are built top down. We first understand the problem, design and then build it.
Learning from data requires a bottoms up approach.
ML systems are different from traditional products. ML systems depend on data. These require regular retraining on data and humans in a loop. Even design and code need to be changes if data pattern changes
The key to build an ML product is - get data, determine what is feasible from data, apply ML with less data. To build generic products, add a variety of data. Adding a variety of data leads to a plurality of methods required to use the data.
Author: Prashant K Dhingra