In this article
Choosing what to work on
What is research taste?
Research taste is the judgement that picks problems worth working on and the experiments that would settle them. Interviewers look for a clear opinion, a reason, a named risk and a cheap first test. It is not a list of fashionable topics; it shows in how you rank problems and how you say you could be wrong.
Why it is asked: the whole round is built to sample this judgement. Read the full explanation: research scientist interview guide.
What would you like to work on if you join this lab?
Name one problem, say why it matters now, and say what a first experiment would be. Tie it to the lab's published direction without pretending to know its internal plans. A specific, bounded answer such as "I would test whether X holds at larger scale" beats a list of topics. Admit what you would need to learn first.
Why it is asked: it shows whether you can pick a direction and defend it. Read the full explanation: research scientist interview guide. Brainstorm-style versions of this prompt are covered in how research brainstorm and problem-framing rounds work.
Which machine learning research topic interests you most, and why?
Pick a topic you can discuss for ten minutes, ideally one you have worked on or reproduced. State the open question, the best current approach, where it falls short and what you would try next. Interest is shown by knowing the weak points of the field, not by reciting a paper's abstract.
Why it is asked: depth in one area is the quickest signal of how you think. Read the full explanation: research scientist interview guide. For framing a vague topic into a question, see research problem-framing rounds.
How do you decide which problem is worth months of work?
Weigh three things: how much the answer would change what others do, whether you have a realistic path to a result, and whether a small experiment can tell you early that you are wrong. A problem that is important but has no cheap first test is risky; say how you would shrink it.
Why it is asked: labs have limited compute and time, and taste decides how it is spent. Read the full explanation: research scientist interview guide. The brainstorm round walkthrough shows how to narrow a big problem live.
What are the most important open problems in AI right now?
Give two or three, not a survey, and justify the ranking. Common candidates are reliable reasoning, understanding what models are doing internally, supervising systems that are smarter than their overseers, and measuring capabilities well. Say which one you would pick and what evidence would change your mind.
Why it is asked: it tests breadth, ranking and honesty about uncertainty. Read the full explanation: research scientist interview guide.
How do you show research taste in an answer?
Make claims you could be wrong about, and say how you would find out. Taste shows as a clear opinion, a reason, a named risk and a cheap test. Vague enthusiasm and long lists show the opposite. Give the strongest objection to your own view before the interviewer does.
Why it is asked: interviewers listen for judgement, not memorised content. Read the full explanation: research scientist interview guide.
Scaling and timelines
How should scaling laws shape model development plans?
A scaling law is an empirical relationship showing how a model's loss changes as model size, data and compute grow. Kaplan et al. (2020) found smooth, predictable trends, which lets a lab forecast results from small runs. Name the limit: a law fitted on past runs may not hold outside the tested range. The models in these studies are transformers, so know that architecture first.
Why it is asked: it checks whether you can use empirical trends for planning without overclaiming. Read the full explanation: research scientist interview guide.
Does the Chinchilla result mean only bigger models matter?
No. Kaplan et al. (2020) had favoured training "very large models on a relatively modest amount of data". Hoffmann et al. (2022) revised that allocation: for a fixed compute budget, scale model size and training data equally. The lesson is balance and compute efficiency, not "bigger is always better".
Why it is asked: it separates a memorised slogan from the actual finding. Read the full explanation: research scientist interview guide.
Are emergent abilities sudden and unpredictable?
Not necessarily. Schaeffer et al. (2023) argue that apparent sudden jumps can depend on the metric chosen: an all-or-nothing score can hide smooth improvement underneath. A careful answer says both views: some abilities may still be hard to forecast, but metric choice must be ruled out first. The deep learning guide covers the training background.
Why it is asked: it tests whether you question a popular claim before repeating it. Read the full explanation: research scientist interview guide.
What if AGI timelines slip and it takes many more years than expected?
One common definition of artificial general intelligence (AGI) is a system matching human ability across most tasks; definitions vary, so state yours. A strong answer separates what changes from what does not: nearer-term work such as evaluation, interpretability and deployment safety stays valuable, while urgency-driven bets lose value. Say which of your projects survive the slower timeline and why.
Why it is asked: it checks whether your priorities rest on one forecast. Read the full explanation: research scientist interview guide.
Knowing it is half the job. Saying it is the other half.
Interview Clozo is an AI interview copilot for real interview rounds: it hears the question and puts a strong answer on your screen, and you say it in your own words. See the features or how it works.
Start Free Trial →Alignment and safety
Which alignment problem is furthest from solved?
Pick one and defend it. A well-supported choice is scalable oversight: how do you train and verify behaviour when the outputs are hard for people to check and the evaluator is weaker than the system? Explain why current methods fall short and what partial progress would look like.
Why it is asked: it shows whether you understand the field's real bottlenecks. Read the full explanation: research scientist interview guide.
How would you test a model for deceptive alignment?
Deceptive alignment is a model behaving well only while it is evaluated. A test: train a small model with a known trigger that switches its behaviour, apply standard safety training, and compare triggered behaviour with an untriggered control. Hubinger et al. (2024) found the backdoored behaviour persisted, and adversarial training can "better recognize their backdoor triggers, effectively hiding the unsafe behavior". These models are deliberately built, not deception that arose on its own, so say what the setup cannot show.
Why it is asked: it tests experimental design on a hard, ill-defined topic. Read the full explanation: research scientist interview guide. Turning a vague worry into a testable question is the skill drilled in research brainstorm and problem-framing rounds.
What are the strongest arguments against Constitutional AI?
Constitutional AI is a method that trains an assistant using AI feedback guided by a written set of principles (Bai et al., 2022). Criticisms to raise: who chooses the principles, reliance on a model's own judgement, and unclear generalisation outside tested cases. Then answer each: publish the principles, audit the judge, test beyond the training distribution.
Why it is asked: it checks that you can criticise a method fairly and then defend or repair it. Read the full explanation: research scientist interview guide.
Does Constitutional AI remove human input?
No. Bai et al. (2022) describe training where "the only human oversight is provided through a list of rules or principles". Humans still write those principles and choose how to evaluate results. The method changes where human effort goes, from labelling many outputs to writing and revising the rules. Say that clearly rather than calling it fully automatic.
Why it is asked: it tests whether you correct a common misreading. Read the full explanation: research scientist interview guide.
At what capability thresholds should extra safety measures trigger?
A responsible scaling policy ties required safeguards to measured capability levels. As of October 2026, Anthropic's Responsible Scaling Policy describes AI Safety Levels, tiers of required safeguards linked to capability thresholds. Propose thresholds based on concrete tests, with a margin before the line, and refer to the current published version rather than recalling details.
Why it is asked: it tests how you turn a safety principle into a measurable rule. Read the full explanation: research scientist interview guide.
Method and conduct
How should you design the experiment you propose?
State the question, the hypothesis, the baseline, the metric, and the result that would prove you wrong. Start with the smallest version that could fail. Name the confounds you can foresee and how you would control them. Do not describe a large programme before the first test. If the experiment needs a training or serving pipeline, the machine learning system design guide covers scoping that infrastructure.
Why it is asked: it shows whether your ideas can become evidence. Read the full explanation: research scientist interview guide. The same habit is tested in problem-framing rounds for researchers.
What do you do when the interviewer pushes back hard?
Treat the challenge as part of the exercise. Restate the objection in your own words, concede what is right, and defend what the evidence supports. Candidates have reported research rounds that were confrontational but respectful. Changing your view for a good reason reads well; holding it for no reason does not.
Why it is asked: research work involves disagreement, and the round samples it. Read the full explanation: research scientist interview guide.
What if you do not know the answer?
Say so, then reason aloud from what you do know: the nearest result you are sure of, a guess, and how you would check it. A reasoned "I am not sure, but I would test it like this" is stronger than a confident invention.
Why it is asked: labs value calibrated honesty about uncertainty. Read the full explanation: research scientist interview guide.
Follow-up questions
How do you evaluate capabilities you haven't imagined yet?
You cannot test for a named capability you have not thought of, so broaden the evaluation: use many diverse tasks, open-ended probing by people trying to find failures, and tracking of behaviour as scale grows. Say plainly that this reduces the risk but does not remove it.
What if capabilities emerge unexpectedly between evaluations?
Run evaluations often, including at intermediate points of training, and leave a safety margin below any trigger threshold. Check whether the "jump" depends on the metric, as Schaeffer et al. (2023) argue it can. Have a response plan ready for a surprise.
How long should a research taste answer be?
A workable target is about one to two minutes for the first answer: the claim, the reason, the risk and the first test. Leave room for the interviewer to push; the follow-up is where the depth is judged. The research interview guide covers the whole round.
- Research taste means choosing problems well and saying how you would test them.
- Give an opinion, a reason, a risk and a cheap first experiment.
- Know the findings behind scaling laws, Chinchilla and emergence, not only the slogans.
- Criticise a method fairly, then answer your own criticism.
- Say what you do not know and how you would find out.
Sources
- Constitutional AI: Harmlessness from AI Feedback (arXiv 2212.08073) — arXiv (Bai et al., Anthropic), 15 Dec 2022
- Scaling Laws for Neural Language Models (arXiv 2001.08361) — arXiv (Kaplan et al.), 23 Jan 2020
- Training Compute-Optimal Large Language Models (arXiv 2203.15556) — arXiv (Hoffmann et al.), 29 Mar 2022
- Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training (arXiv 2401.05566) — arXiv (Hubinger et al.), 10 Jan 2024
- Are Emergent Abilities of Large Language Models a Mirage? (arXiv 2304.15004) — arXiv (Schaeffer et al.), 28 Apr 2023
- Responsible Scaling Policy (Anthropic; page read 2026-10-26) — Anthropic, 26 Oct 2026
