MAY - JUNE 2023CIOAPPLICATIONSEUROPE.COM8IN MYOPINIONo understand applied AI these days, you have to think back to the time before the Industrial Revolution. Back then, if you wanted an iron plough, you had to ask the local blacksmith, an expert craftsman with years of training and experience. He would create a plough just for you.But with the advent of the factories, the majority of the ploughs were made by machines or on assembly lines. Designing those machines and assembly lines required an even higher level of skill, but the factory workers needed much less expertise than the blacksmith to produce the plough. The level of skill needed to create the plough varied depending on where a worker was in the chain.How does this apply to AI?Similarly, several years ago, creating a machine learning model required intricate craftsmanship. But now, more and more problems have out-of-the-box solutions that are, at the very least, `good enough'. For example, for making predictions from tabular data, gradient boosting approaches, such as XGBoost, LightGBM, and CatBoost, have clearly won. Similarly, for working with natural language, there has been an explosion of models such as BeRT, now easily available through HuggingFace libraries. Standard solutions for time series forecasting have lagged (with Facebook's Prophet being the exception), but lately there has been a wave of big companies such as Uber, LinkedIn and Facebook open sourcing their approaches. We've also seen a rise in `easy to use wrappers' for academic work, such as PyTorch Forecasting.Some domains are still not fully commoditised, such as natural language text generation (including chatbots) or image understanding and generation. Any packages that do exist here are not high profile, but there's been huge progress already and it's likely we'll see these appear in a year or two.Considering all of the above, it seems likely that in time, the majority of your data scientists will not be craftsmen, but assembly line workers. You still need some in-house experts to design that assembly line, but the rest should be able to cross-train from developers or analysts, or be hired from the large number of people attending machine learning courses. It's very possible that in three or four years, machine learning will turn from a job description to just another skillset, like SQL or Excel macros. Can you rely solely on `out-of-the-box' solutions today?In short, no. There's still a lot of work to be done to advance the evolution of AI, and currently not enough machine learning projects are making it into production. A significant cause of that is the lack of maturity in MLOps - which is the software stack needed to deploy, use, and monitor models in production. Unlike the LAMP stack or the ELK stack in other domains, there is no single standard stack for that, and there are dozens of providers for specific components or component combinations of that stack, vying for supremacy. In short, there's no set way to put these solutions into production, and that makes them hard to introduce, market, and implement.THE COMMODIFICATION OF AI: WHAT IT MEANS FOR YOUR BUSINESS, AND HOW TO CAPITALISE ON ITEGOR KRAEV, HEAD OF AI, WISETEgor Kraev
< Page 7 | Page 9 >