Practiced one out loud? Paste your answer into the free self-check and get a score in 30 seconds
What the interviewer is testing
Whether you know how reasoning models are trained and where they help, and whether you can make a cost-aware product decision rather than defaulting to the strongest model.
Answer framework
- Training: reasoning models are trained, typically with reinforcement learning on tasks with verifiable answers, to think through a long chain of thought before answering.
- Inference: they spend more compute per query, so they are slower and cost more.
- Strengths: math, code, multi-step planning and problems where checking intermediate steps matters.
- Weaknesses: poor value for simple lookups, short rewrites and style tasks.
- Product design: route only hard queries to them, set a reasoning budget, and evaluate cost per solved task.
Sample answer
A standard chat model answers more or less directly after reading the prompt. A reasoning model is trained, typically with reinforcement learning on problems where the answer can be checked, such as math and unit-tested code, to produce a long internal chain of thought first. It learns to break problems down, try approaches and check its own work, which is why it does better on multi-step problems.
The cost is inference-time compute. The hidden reasoning can be many times longer than the visible answer, so latency and price per query go up. For a quick factual lookup or rephrasing an email, that extra thinking adds little.
In a product I would not send everything to a reasoning model. I would route by difficulty: a cheap classifier or rules send hard analytical or coding requests to the reasoning model with a capped reasoning budget, and everything else to a fast chat model. Then I would measure accuracy, latency and cost per successful task on both paths.
Likely follow-ups
Common mistakes
- Using a reasoning model for every request regardless of cost and latency.
- Assuming the visible reasoning is a faithful record of how the answer was produced.
- Ignoring time to first token in user-facing flows.
- Comparing models on accuracy only, not cost per solved task.
The 0 to 1 AI Engineer Interview Playbook covers the answer frameworks, system-design walkthroughs and an 8-week plan.
Get the playbook 40% off with code EARLYBF until Nov 3, 11:59 PM PTEARLYBF: 40% off until Nov 3Smaller start: Score your answers against 50 rubrics, $9. 50 AI engineer questions, one rubric each, PDF on Gumroad.
Interviewing for PM or TPM? See the full PM package, $59 on Gumroad: a resume system, a PM, TPM and PGM interview handbook, a 1:1 cheat sheet and a PM playbook. PM and TPM content, not AI engineer prep.
Related questions
Based on question #13 in the full question bank.