The Machine Learning career path

Research and applied are different jobs, and the pay gap between them is not the interesting part.

By Surojit Chakraverti — ex-Citi and Rothschild M&A, now a hedge fund CIO

Machine learning splits early into research and applied work, and the distinction is not seniority. A research scientist is trying to make a method work that does not yet; an applied engineer is making a method that works survive contact with production traffic.

The field also moves faster than the ladders describing it, so titles travel poorly between firms. Read the work, not the level.

The ladder

What changes at each level, rather than what the title is.

01
Applied ML engineer or junior researcher
Years 0 to 3
Running experiments someone else framed, and building the pipelines around them. Most of the job is data and evaluation rather than modelling, which surprises people arriving from coursework.
02
Senior applied, or research scientist
Years 3 to 7
Framing the problem rather than the experiment. You decide what would count as working, which is the harder half, and you own the trade-off between a better metric and a shippable system.
03
Staff, or senior research scientist
Years 7 to 11
Direction. Choosing which bets the group takes and killing the ones that are not paying, across several teams.
04
Principal or research lead
Year 11 onward
Agenda-setting, hiring, and defending a programme internally over horizons longer than a planning cycle.

The hours

Similar to software engineering, forty to fifty, with the notable exception of research groups working to conference deadlines, where the weeks before a submission look like a different job.

Compute is the real constraint rather than time. Waiting on a training run shapes the working day more than anything on a calendar.

What it pays, by level

Base and bonus at every rung, aggregated from published surveys rather than from anecdote.

Machine Learning compensation, level by level

Where people go next

Quantitative research
Funds hire directly from this pool and pay above technology rates. The methods transfer; the adversarial setting does not, and that is the adjustment.
Standard software engineering
Common and undramatic, often for scope or stability rather than pay.
Founding
The most active area for it in years, and the reason experienced researchers are hard to retain.
Academia
The reverse of the usual flow, and generally a decision about what you want to work on rather than about money.

Open machine learning roles now

Every tracked programme, updated daily.

Common questions

Research scientist or applied ML engineer: which should I aim for?

Applied if you want your work in front of users and a broader set of employers, research if you are willing to trade that for depth on unsolved problems. Research roles at the top labs are far fewer and usually expect published work.

Do you need a PhD for machine learning?

For research at the leading labs, generally yes or an equivalent publication record. For applied engineering, no, and a strong software background plus demonstrable ML work is the more common route.

Does machine learning pay more than software engineering?

At the same level and firm, modestly, and the gap is widest in research seats at the labs competing hardest for them. The larger determinant is still the employer rather than the specialism.

Deciding is one thing. Getting in is another.

This page is the map. The prep track is the route: what the interviews test, what firms look for, and the questions you will actually be asked.

Machine Learning interview prep Book a diagnostic

Guides for this path

Data Science Interview Questions in Finance