A recent episode of the MLOPS Community Podcast discussed the myths of machine learning and data science work.
Host Demetrios Brinkmann interviews data scientist Nikhil Suresh. He runs a blog where he gives honest opinions on technology trends. The interview is based on one of his recent blog posts on AI hype.
I recommend listening to podcasts and reading blog posts both (Suresh uses many profane languages, but I don't think this needs to prove your point).
But here are some of the distilled facts I have completely agreed to, followed by some of my own observations about the larger picture, which may be missing argued.
Machine learning and data science work is far less than it is advertised. Suresh writes: “The number of companies launching AI initiatives was far greater than the number of actual use cases. Most of the market was merely glyfters and incompetent people (sometimes both!) and were able to harness the hype to inflate personnel, promote and view them as thought leaders.”
Given the hype surrounding artificial intelligence and machine learning, all companies want to brand themselves, such as “AI-First” or “AI-focused.” Over the past few years, there have been many ads about the sexiest jobs of the 21st century: data scientists and machine learning engineers.
However, the actual work of ML engineers and data scientists is very different from frontier research, usually done in deep learning and generation AI. Most of the work is basically creating simple models such as xgboost. Deep learning is rarely necessary.
Without a doubt, these works have a lot of value, but they are not advertised and do not require huge data science skills. This means that with basic machine learning skills and a bit of work experience you can probably step into the door, but others can do so too.
“I have gradually come to the conclusion that the market for specialist data scientists is actually relatively small in terms of the number of use cases,” Suresh said on the podcast. “But the hype in the market really has created an over-age. There are many roles, but only some of those roles are real.”
The infrastructure is not ready for data science and machine learning projects. According to a blog post from Suresh, “The data science work began to evaporate, and the hype cycle has moved from all AI initiatives that have failed to make progress and started inch towards data engineering.”
This also aligns with my own experiences and conversations with other practitioners in the field. Data science requires special arrangements between data lakes and warehouses. However, most companies prepare their data infrastructure for classic business intelligence and analytical operations. Coordinating them for data science projects requires extensive restructuring. Therefore, in many cases you sign up for data science or machine learning roles, you need to do a lot of data engineering work to create the right data pipeline to train your models.
As mentioned in the podcast, “Data engineering comes before AI.”
There is a gap between business and machine learning teams. “Businesses usually don't understand how machine learning interventions interact with businesses,” the podcast discussed. “They usually want rows of graphs and don't know what to do with the predictions the model has made.”
Again, this is a great thing. One of the issues I encountered regularly was the gap between the machine learning team and the business team. Business experts don't know what to expect from machine learning models. Therefore, the ML team does not understand the business needs and how the model can make changes. One of the things I advocated is to strive to create a collaborative business and machine learning team that can work backwards from business needs and interventions to forecasting, data requirements and models. The BIZML framework by Eric Siegel provides a great starting point for this.
Let experts handle AI technology. Suresh said, “Unless you're one of a handful of companies that know exactly what uses AI, AI is not necessary for anything. Rather, you don't have to do anything to enjoy the benefits.
Being AI-First means doing AI at the end. You don't need to think of AI as an absolute necessity for everything in your business. You need to think about the product and solve the customer's problems. AI is sometimes a useful feature. And today, most solutions you use for your business have AI capabilities whenever you need them.
Focus on solving the problem. Use the tools available whenever possible. Create your own machine learning models and products only if they are not yet available elsewhere.
The whole picture
Subscribe to continuous reads
Become a paid subscriber to access the rest of this post and other exclusive content.
