ML Education
at Google
When training becomes transformation
The problem
In 2016, Google needed to become an ML-first company. That was easier said than done, even for Google in 2016. The infrastructure, culture, and capabilities all had to shift simultaneously—across 100,000 people and hundreds of product teams.
Top-down mandates and corporate training fail because they don't change how people actually think and work. Bottom-up adoption is too slow when you're racing competitors. And the classic approach—train a core team of experts who gradually spread knowledge—creates bottlenecks that last for years.
The real impossibility: How do you transform an organization's fundamental capabilities when the infrastructure to support those capabilities doesn't yet fully exist?
The insight
Education at scale creates irresistible organizational pressure. When you announce that you're training the entire company on something, leadership must ensure the infrastructure exists to support what you're teaching. Otherwise, they look incompetent or the training fails publicly.
This makes education a strategic lever, not just a knowledge transfer mechanism. Done right, it forces alignment across tooling, infrastructure, and decision-making frameworks—because those things have to work for the training to succeed.
My team had seen this pattern work before. When Google needed to become a mobile-first company, we launched Android Bootcamp—company-wide training that forced the infrastructure development the company needed. Before it, the Android team owned the mobile experience for every Google product. After it, even Apple featured Google apps as examples of excellence on iOS.
The curriculum needed to teach engineers
how to think about ML, not just
what to know about it.
What we built
Machine Learning Crash Course (MLCC) was designed to work at multiple levels simultaneously.
For the entire organization: ML became a cultural primitive—an idea "in the air", as a way of thinking about problems, even for people who had never (and would never) write an ML algorithm.
For the engineering organization: Product managers, UX designers, writers, and engineers could have informed conversations about when ML was appropriate, asking questions like:
- Is there a clear, repeatable decision to be made frequently enough to warrant automation?
- Can we collect or generate enough examples?
- What's our tolerance for error?
- How would we validate automated decisions?
- Is there a simpler, mechanical solution?
For engineers specifically: Hands-on capability to implement ML systems, including the mindset shift from "I have an algorithm that transforms data" to "I can collect data that trains an algorithm."
Because of Google's structure, anyone from any background could take the course. Over time, we expanded to serve technical experts on specific topics and non-technical audiences on strategic thinking.
The results: We trained more than 40% of Googlers and made the Machine Learning Crash Course publicly available, where it reached over 4 million people per year worldwide.
How it propagated
The measure of cultural transformation isn't course completion, it's behavior change. Internally, we tracked ML pipeline creation through to production: were engineers actually building ML systems, or just learning about them? The trend was positive, and small ML-based features began appearing across most Google products—not as flagship launches, but as quiet evidence that the mindset shift had taken hold.
The content was strong enough that its value became obvious beyond Google. We made the MLCC publicly available, where it reached over 4 million people per year, worldwide. Google Cloud recognized a commercial opportunity: bringing select, high-value customers on-site to receive the same training Googlers received, building ML systems alongside Google engineers. Training became a cornerstone of a new sales pipeline.
The furthest reach was the squishiest to measure: MLCC content ultimately reached key opinion formers in policy and government, elevating the quality of public discourse around ML and AI at a moment when that discourse badly needed elevation.
We had built something to transform Google. It ended up contributing to how the world thinks about machine learning.
What it demonstrates
This work required seeing organizational transformation at a different level than most people operate. The real work was understanding that education at scale creates pressure that forces infrastructure, tooling, and culture to align. Building great training content was a tool, not an outcome.
That insight only matters if you can execute: leading a team of nearly 30 people across 10 global offices, coordinating with business units for tailored delivery, and managing quality at a scale where millions of people are watching.
It's work well-suited to high-stakes organizational challenges: when the conventional approaches (mandates, expert teams, incremental change) won't move fast enough, and you need a forcing function that makes transformation inevitable rather than aspirational.