Read the opening of chapter 1
The opening of chapter 1, exactly as printed.
Chapter 1: Judgment at the Top: What Separates Competent from Excellent
Two engineers with the same title, the same three years of post-senior experience, and functionally the same skill set compare offers over a shared Google Sheet. One number reads $172,000 base with a modest bonus. The other reads $310,000 in total comp once a four-year equity grant is annualized. Neither person is being lied to. Neither recruiter is wrong. The gap isn't a negotiating failure on one side; it's two different markets wearing the same job title, and until you can explain exactly why that gap exists, you're not ready to negotiate in either of them.
That's the actual condition of the market you're operating in at this level, and it's the reason this book starts here instead of with a tools chapter. Below senior, the job is mostly legible: master the loop, ship models, learn the frameworks, and the market rewards you roughly in proportion to how well you do those things. Above it, the legibility drops out. Compensation surveys disagree with each other by tens of thousands of dollars for the same role. Specialization premiums get reported as a single number by outlets that measured completely different things. The skill that got you promoted to senior — executing well against a known loop — stops being the thing anyone is evaluating you on. What replaces it is judgment: the ability to make a defensible call when the data itself doesn't resolve cleanly, and to know the difference between noise you should ignore and signal you should act on.
You already know the rough shape of the ladder because you've either climbed part of it or watched colleagues do it. It's worth stating plainly anyway, because the numbers matter for what comes next: junior ML engineers land around $110,000 in base salary in most US markets; mid-level engineers sit in a wide band above that; senior engineers cluster between roughly $150,000 and $210,000; and principal or staff-level engineers routinely clear $250,000, with the ceiling determined more by company and equity structure than by title. If you're at this point in your career, you're either at senior now or close enough to it that the next real decision — not the next raise, the next decision — is the one this ladder forces on everybody who reaches it.
Keep readingShow less
That decision is the fork between the hands-on technical track and people management. Somewhere around senior-to-staff, the ladder stops being one ladder. One branch keeps you writing code, owning systems, and going deeper technically: staff engineer, principal engineer, distinguished engineer. The other branch has you managing engineers instead of systems: engineering manager, director, VP. Both branches can reach the same comp ceiling. Neither is a demotion or a consolation prize, and neither is objectively "more senior" than the other; they're different jobs that happen to share a level number. What actually differs is what you get evaluated on. The technical track keeps grading you on the quality of your judgment about systems — architecture calls, build-vs-buy calls, when to accept technical debt and when to pay it down. The management track grades you on judgment about people and organizations: who to hire, how to resolve a disagreement between two senior ICs who both have a legitimate point, when to protect your team from a deadline that shouldn't exist.
If you're near this fork and haven't consciously chosen, that's worth noticing as a fact about your career, not a personal failing. A lot of people drift into management because it's offered, not because they evaluated it against the alternative. The rest of this book is written primarily for the technical track, because that's where deep ML-specific judgment keeps compounding as a skill. But the fork itself is the first real judgment call at this level, and pretending it isn't a fork — treating management as just "senior, but with reports" — is exactly the kind of false precision this chapter is going to keep warning you about.
The honest test for which side of the fork you belong on isn't "do I like people" or "do I still enjoy writing code" — most senior engineers answer yes to both and that tells you nothing. The sharper question is where your best hours actually go when nobody's forcing a choice: toward untangling why a model's offline metrics and production behavior have quietly diverged, or toward untangling why two capable engineers on your team have dug into incompatible positions about the same design and need someone to actually resolve it, not just referee it. Both are real judgment work. Neither is easier than the other, and staff engineers who assume management is the "softer" track discover the opposite fast once they're the one accountable for a hire that didn't work out. What should worry you more than picking the "wrong" track is picking either one by default, because the skills you build in your first two years past this fork are hard to un-build, and switching back later costs real time.
Once you've internalized that the premium ranges above are ranges and not points, the next skill is knowing what else is hiding inside a quoted number besides the baseline it's measured against. Cash-versus-equity mix is one of the biggest hidden variables: a "40% premium" quoted at a public company paying mostly cash means something structurally different from the same headline number at a pre-IPO startup paying it mostly in options that may be worth far more or far less by the time they vest. A recruiter citing a premium rarely volunteers which mix they mean, and you're entitled to ask directly. "Is that number mostly cash, or does it assume the equity hits its current valuation" is a fair, normal question at this level, and how a recruiter answers it tells you almost as much about the offer's quality as the number itself.
Here's where the false-precision habit actually costs you money if you don't break it. Ask five sources what an MLOps specialization is worth over a generalist ML engineer salary, and you'll get four answers that cluster in the same neighborhood and one that doesn't. The cluster sits at roughly 25% to 40% above baseline, consistently enough across methodologies that you can treat it as a real, load-bearing number. Ask the same five sources about an LLM or GenAI specialization premium, and the answers scatter from about 15% up to 60%, and none of them are lying to you either. They're measuring different things: different baseline comparison groups, different seniority bands folded into "LLM engineer" as a category, different treatment of equity versus cash, different geographic mixes in the sample. A recruiter who tells you "GenAI engineers get 40% more" isn't giving you a fact about the market. They're giving you one point sampled from a distribution wide enough to swallow both a real premium and no premium at all, depending on where in it you land.
