Machine learning is brilliant but opaque, and a generator of technical debt. The concept of ML models as “black boxes” is very visible in the media now. Perhaps it was a little less so when I gave the talk Acceleration With A Steering Wheel at DevOps 2018 in Singapore.

Proposal to rename machine learning to “racist linear algebra” – @hillelogram

Technical debt is not the only bad consequence of unmanaged models. Tech debt gets blamed for many things today, not without reason, but not necessarily with a lot of technical precision. I framed ML as an accelerant, which requires new elements of control to use effectively. Jet fuel is another accelerant, and it behaves differently in a working jet engine versus being poured all over your living room. In a jet engine, there are social and mechanical forms of control which make the system work. And DevOps techniques are about gaining control over your software process.

This is not central management control! Control can’t be a bottleneck on everyday decisions. On the contrary, it’s only by giving individual team members control over quality that makes Agile and DevOps techniques work. The ability to make frequent, relevant adjustments allows problems to be caught through early feedback, and engineered out throughout the process. Therefore: ML as accelerant, DevOps as steering wheel.

Slides here (minus a little technical detail from my former employer). I notice Gregor Hohpe, in “great minds think alike” fashion, made the same analogy recently.