The Future of MLOps: and how we got here
Most AI/ML projects start shipping models into production, where they can deliver business value, using the no-process process. That is, people just do their best by creating an ad-hoc process with familiar tools. This works for tiny teams at first, but as the team grows you’ll discover significant chaos and pain trying to operationalize AI.
We’ve been here before. Software development in the 90s was a lot like the “no-process process” for ML today. And just as the paradigm shift known as DevOps brought reproducibility, collaboration and continuous delivery to software. Applying the same principles to ML can bring the same benefits to AI projects. Without it, AI projects fail and create financial and reputational risk.
In this talk, Luke will dive deep into the important differences between software development and ML which mean you can’t just reuse DevOps tools. He will share his team’s research comparing the evolution of Software Development & DevOps with that of Machine Learning. Luke will present a manifesto, and propose an architecture and a set of open source tools to make ML reproducible, accountable, collaborative and continuous.