LIKELY QUESTIONS
- Walk me through the most impactful generative AI project you have led end-to-end, including the research hypothesis, training setup, evaluation methodology, and production outcome.
- Midjourney cares about image, video, and world models. Which of these areas are you strongest in, and what original ideas would you explore in your first 6 months here?
- How have you trained and optimized large-scale models under real-world compute, data, and iteration constraints?
- What is your approach to improving diffusion or transformer-based generative models beyond incremental benchmark gains?
- Tell us about a time you invented or adapted a novel architecture, loss, sampling method, or training strategy that materially improved model quality or efficiency.
- How do you evaluate generative models when standard metrics do not fully capture creativity, aesthetic quality, controllability, or user value?
- In a small elite research team, how do you balance open-ended exploration with shipping useful systems quickly?
- If you joined Midjourney tomorrow, how would you prioritize research across image generation, video consistency, 3D or world understanding, and deployment readiness?

BEHAVIOURAL QUESTIONS
- Tell me about a time you pursued a risky research direction that others were unsure about.
Model approach: Situation: high-stakes project with a plateau in model quality. Task: identify a new direction with limited evidence. Action: formed a clear hypothesis, built a lightweight validation plan, defined kill criteria, ran fast experiments, communicated tradeoffs, and pivoted details while defending the core idea. Result: achieved measurable gain in quality, efficiency, or capability, and showed disciplined risk-taking rather than intuition alone.
- Describe a time you had to work with significant autonomy and little structure.
Model approach: Situation: ambiguous research goal in a fast-moving environment. Task: create clarity, milestones, and momentum without heavy management. Action: translated the broad goal into concrete experiments, set success metrics, built a decision log, coordinated with a small team, and regularly surfaced findings. Result: delivered a research outcome or shipped prototype quickly while showing ownership and judgment.
- Tell me about a disagreement with another strong technical researcher and how you handled it.
Model approach: Situation: conflict over architecture choice, evaluation method, or scaling plan. Task: resolve the disagreement without slowing the team. Action: reframed the debate around hypotheses and decision criteria, designed an apples-to-apples experiment, stayed open to being wrong, and kept communication respectful and data-driven. Result: reached a better technical decision, preserved trust, and improved team rigor.
- Give an example of taking a project from experimentation to deployment.
Model approach: Situation: promising research result needed to become a usable system. Task: bridge research code and production constraints. Action: hardened the pipeline, improved inference efficiency, addressed reliability and quality regressions, aligned with stakeholders on launch criteria, and monitored post-launch behavior. Result: successful deployment with measurable adoption, performance, or quality improvement.

SMART QUESTIONS TO ASK
- What are the hardest unsolved research problems the team is prioritizing right now in image, video, and world models?
- How do you evaluate success for researchers here: breakthrough novelty, model quality, product impact, training efficiency, or speed of iteration?
- What does the research-to-product loop look like at Midjourney, and how often do researchers directly influence deployed systems?
- How does the team make compute allocation decisions across exploratory work, scaling experiments, and product-driven improvements?
- What distinguishes researchers who thrive at Midjourney from those who are strong technically but not the right fit for this environment?

RED FLAGS TO WATCH FOR
- Vague answers about evaluation, safety, or model quality criteria, which can signal unclear research priorities or weak decision-making discipline.
- Overemphasis on nonstop heroics with little discussion of sustainable execution, collaboration, or research rigor in a small team.
- No clear explanation of compute access, data strategy, or how ideas move from experimentation to deployment, suggesting execution bottlenecks.