Avoid Job‑Hunting Pitfalls: How a NLP PhD Secured an OpenAI Offer After 57 Interviews

Background

Alisa Liu, about to graduate from the University of Washington with a PhD in NLP, announced on X that she will join OpenAI. She wrote a detailed blog to document the entire job‑search process, hoping to reduce confusion for future candidates.

Interview Volume

During a six‑year PhD, she applied for Research Scientist and Member of Technical Staff roles at 11 companies, completed 57 formal interviews, received 46 recruiter calls, and held 16 deep‑dive conversations after receiving offers. She also engaged in countless informal networking chats.

Types of Interviews

She categorised the interviews she experienced into seven types:

ML coding questions – implementing classic architectures (e.g., Transformer), decoding strategies (e.g., beam search), or creative variants; PyTorch is essential, occasional numpy‑only tasks appear.

General algorithm questions – LeetCode‑style problems, often with a background story; the underlying logic overlaps with ML coding.

Technical discussion (no coding) – deep‑thinking questions where the interviewer presents a research goal, asks you to design experiments, and probes your reasoning; also rapid‑fire knowledge‑breadth questions.

Research experience discussion – presenting a past project, explaining why the topic was chosen, unique judgments made, and future directions; tailored to each company’s focus.

Behavioral interview – the only type where she recorded a clear failure; she ran out of concrete stories and could not answer basic questions, learning to map PhD stories to common behavioral frameworks in advance.

Mathematics interview – ranging from logical puzzles to paper‑and‑pencil calculations; she recommends reviewing probability, linear algebra, and calculus.

Job talk – a concise, focused presentation (her talk centered on tokenization) that weaves together a first‑author paper, co‑authored works, and ongoing projects.

Preparation Strategies

She emphasizes that there are no shortcuts. Her preparation included:

Creating and continuously updating an LLM note‑taking notebook and a math notebook for specific interview needs.

Studying Stanford’s “Language Modeling from Scratch” course to build a coherent knowledge map.

Deep‑diving each concept by reading blogs and papers, conversing with AI experts, and implementing from scratch.

Practising Transformer implementation until it becomes muscle memory, as it appears frequently in interviews.

Simulating real interview conditions by disabling all AI assistance.

Customising study for each interview based on the job description, the company’s technical direction, recruiter hints, and public reputation.

She also shares a costly lesson: sleeping only two hours before a technical interview led to a ten‑minute stall on a simple off‑by‑one error, underscoring the importance of adequate rest.

Offer and Negotiation

After receiving offers, she entered a prolonged phase of deep conversations with future teammates and managers, company meals, and extensive email exchanges. Negotiation was the most challenging part because PhDs rarely practice it. She notes that initial offers usually contain negotiation space; recruiters often say they do not expect you to accept the first offer. Her approach was to write down what she would disclose, what she would keep private, and rehearse key talking points before each recruiter call.

Emotional Challenges

She describes managing intense emotions throughout the months, feeling pressure from peers, and making high‑impact decisions with incomplete information. The process left her on the brink of burnout, but also boosted confidence and reduced anxiety in academic discussions. She reflects that the PhD journey is about generating and executing good ideas, not merely worrying about livelihood.