The Barbell Market for ML Engineers
What four years inside Google taught me about where the profession is heading
I’ve watched the job of a Machine Learning Engineer change more in the last two years than in the previous decade combined.
And it’s creating two very different winners while quietly eliminating everyone in between.
I call it the barbell market. Heavy on both ends. Nothing in the middle.
The Barbell
In finance, a barbell strategy means you put weight at the extremes and avoid the middle entirely. Maximum risk on one end, maximum safety on the other, nothing in the comfortable center that feels safe but isn’t.
That’s exactly what’s happening to ML engineering right now.
On one end of the barbell: the people building foundation models.
On the other: the AI engineers shipping products on top of them.
Both ends are thriving. Both ends have real demand, real leverage, real futures.
The middle, where most ML engineers actually live, is the bar itself. And it’s getting crushed.
One End: The Foundation Model Engineers
At one extreme, you have the engineers and researchers building the actual models.
The LLMs.
The generative recommendation systems.
The E2E architectures replacing stacks that took entire teams years to assemble.
These are the people working on pre-training runs, RLHF pipelines, model architecture at a scale where the problems are genuinely unsolved.
They’re not applying models. They’re building the substrate everyone else builds on top of.
This group is thriving. Demand is extraordinary, compensation reflects it, and the work is getting more technically rich, not less.
The better foundation models get, the more critical it becomes to have people who can push them further. Paradoxically, the AI boom hasn’t commoditized this work: it made it more valuable.
The engineers closest to core model work — training infrastructure, large-scale optimization, architecture research — exist in a completely different labor market than everyone else. The gap widens every quarter.
The catch: this end of the barbell is almost impossible to enter mid-career. You don’t get here with hustle and tutorials. The barrier is genuine depth in a narrow area most people never develop.
The weight is heavy because the bar to lift it is real.
I write about ML systems in production — the tradeoffs, the architecture decisions, the stuff that doesn’t make it into papers. If you want to go deeper, the paid tier covers the technical details I can’t fit in free posts.
