I have been listening to the podcast This Week in Machine Learning and AI for a couple of years now and I always enjoy the many scalable platform shows. So, when the TWIML.con AI Platforms conference was announced last spring I knew it wanted to attend. Even better is was being held in San Francisco, just a 45 mile drive up Interstate 280 from where I work.
When I first started listening to the TWIMLAI podcast I focused on interviews with ML practitioners from Apple, Facebook, Google, LinkedIn, SalesForce and other mega-scale companies talking about the platforms they had built internally. Then I started to listen to more independent software vendors (ISVs) discuss their commercial software products. There were several similarities between how these groups were approaching their designs but also some key differences that I wanted to understand better and I thought the AI Platforms might help me get some clarity.
I wasn’t disappointed. There were speakers from Airbnb, Facebook, LinkedIn, SurveyMonkey, Twitter, Workday giving updates on lessons learned from building platforms for machine learning and data science. This was a relatively small conference and the speakers were easily accessible during breaks and during meals for questions. It wasn’t until a panel session at the end of the first day titled “Scaling ML in the Traditional Enterprise” that I had my first flash of insight. The speakers on the panel first tried to explain what they meant by a “traditional enterprise”. They were talking about organizations whose primary value was derived from something other than a software product. These organizations were not experts in dev/ops or other aspects of the software development process that were so deeply embedded in the DNA of the companies that were born from cloud-native software development.
I often talk about how data science is a “team sport” and since most data modeling ends up embedded in software applications, people with strong software architecture and dev/ops skills are essential to the success of the team. The panelists on the “Scaling ML in the Traditional Enterprise” were clear that in the traditional enterprise as they had defined there was a serious shortage of developers and architects to bring on the team. Data scientists in traditional organizations were mainly looking for commercial software platforms that would allow them to create scalable web services or full-blown apps with UI that could expose their models to the rest of the organizations. Data scientists at the cloud-native software companies were surrounded by a rich ecosystem of people and process that they needed to integrate with as opposed to figuring out how to be a one-stop service on their own. Traditional organizations were trying to make up for a shortage of skilled people by adopting a commercial platform.
My insight at the end of day one was based on a sample of one, but by the end of the conference, I had plenty of corroborating data. The “Team Teardown: Evolving an ML Platform at SurveyMonkey” panel on day two highlighted the other side of the coin. Here was a team of scalable platform experts that understood dev/ops for rapid feature releases that needed to define how the data science team was going to integrate with the best practices they had developed. There was no commercial software platform that could help with these challenges. They needed to develop process and environment agreements that would allow them to be more agile and reliable.
The “Unconference” held at the end of the second day really highlighted the importance of these differences in how different organizations approach ML platform challenges. There were three separate sessions proposed by three different attendees that covered “AI Platforms – Build vs Buy”, “Evaluating AI Platform Vendor Offerings” and “Developing a Dev/Ops Culture for ML”. If you are working in a “traditional enterprise” setting and not getting value from your data science initiatives because of trouble with “last mile execution” getting models into production, you are not alone.
Thanks for reading,
Phil Hummel – Twitter @GotDisk