Chip Huyen on Machine Learning Interviews (Full Stack Deep Learning - November 2019)

Machine learning hiring is less about “perfect” interviews and more about navigating a noisy, expensive, and often inconsistent process—so candidates should optimize for fit, signal quality, and question selection rather than chasing a single formula for success. Chip Huyen, working on productionizing AI research at Nvidia, draws a sharp line between research and engineering roles, then maps how that distinction shows up in recruiting pipelines, resume screening, and interview question design.

A central theme is that “research” and “applied research” differ in time horizon and outcome expectations. Fundamental research targets long-term answers to theoretical questions, while applied research aims for practical solutions with nearer-term commercial impact. Even in applied settings, cutting-edge machine learning remains heavily empirical because many techniques work without a strong theoretical framework. Huyen also distinguishes research scientists (who originate ideas) from research engineers (who scale and operationalize those ideas), noting that modern model scaling makes engineering skill essential for turning prototypes into systems.

She then broadens the career landscape: machine learning jobs can come from multiple starting points—direct ML engineering from bachelor’s or master’s programs, PhD-to-research roles, transitions from data science, and software engineers retraining via coursework or bootcamps. Another pathway is hiring “adjacent” researchers from statistics or physics and training them on the job. The common thread is time: there’s no shortcut to becoming an ML expert in weeks, and online “five easy steps” claims should be treated skeptically.

Huyen emphasizes that big-company machine learning differs from startup machine learning. Large firms can afford expensive compute-heavy research and often hire specialists who focus on narrow components. Startups move faster across domains and typically prefer generalists who can handle multiple moving parts. That difference also shapes how candidates are evaluated and how many interviews are required.

On interviewing, she argues that outcomes depend on variables beyond raw ability. Interviewers often receive little training, may use “pet questions,” and can be mismatched to the candidate’s strengths (for example, a computer-vision expert facing an interviewer who doesn’t know the area). Mood, stress, and day-to-day factors can also skew results. Because of this, a rejection is not a reliable measure of competence, and companies sometimes revisit candidates later.

Recruiting itself is constrained by cost and headcount. Huyen notes that hiring agencies can take 20–30% of first-year salary, and hiring managers under pressure may choose “good enough” candidates whose schedules and signals align with the immediate need. Resume screening is informal and often biased toward recognizable signals—previous employers, GitHub, open-source contributions, awards/papers, referrals, and sometimes school names—while also being limited by recruiters’ lack of engineering background.

Finally, she critiques common interview question types. Easy recall questions are weak predictors of excellence; trick questions, overly specific name/term tests, and open-ended questions become unfair when interviewers expect a single “correct” solution. Better questions probe understanding and assumptions (e.g., how k-means behaves under dataset changes), and can use structured formats like multiple choice, quizzes, code walkthroughs, pair programming, or two-interviewer setups to improve quality.

To ground the discussion, she shares analysis from thousands of Glassdoor reviews across major tech companies, finding that companies with lower on-site offer ratios often have higher offer acceptance rates—suggesting that selectivity and offer strategy interact with candidate behavior. For candidates, her practical advice is to interview when not desperate, run a small number of targeted interviews to calibrate, and treat internships as a more attainable entry point since they’re cheaper for companies and often convert to full-time roles.