AI-native products
Runway
Generative video and image tools where cost per generation and iteration speed are the product.
Overview
Runway builds generative video and image tools, and increasingly frames the underlying work as general world models — systems whose primary input and output is video rather than text. Their public research covers the Gen video model line, world models, character and real-time avatar work, and image generation. Their careers listings mix research scientists and research engineers with backend, API and infrastructure roles, and describe a team of researchers, engineers, artists and designers rather than an engineering org with a design function attached.
The product constraint is unusual and worth internalising. A creative tool is used iteratively: someone generates, dislikes the result, changes one thing and generates again. That loop is what people pay for, so time to first frame and cost per generation are product decisions rather than infrastructure details. Their own API is built around asynchronous task submission and polling, which tells you generations run long enough that the entire interface has to be designed around waiting well.
Quality is also genuinely subjective. No accuracy number settles whether a shot is good, so evaluation leans on structured human preference and on the judgement of the artists working alongside engineers. Candidates who can only defend a model with a benchmark score struggle here; candidates who can talk about controllability and what a user does when the output is nearly right do well.
What Runway visibly values
Inferred from public material, not from a leaked rubric. Treat these as the themes worth emphasising rather than as a scoring sheet.
- Real understanding of video generation — diffusion and autoregressive approaches, temporal consistency, conditioning and control signals, and why long clips are harder than long text.
- Inference cost per generation. Video is expensive to produce; step reduction, distillation, caching, resolution laddering and batching are the difference between a viable feature and one that loses money.
- Iteration-loop design. Preview quality, deterministic seeds, partial regeneration and cancellation matter more than a single flagship result, because the user will run the model twenty times.
- Evaluation without a ground-truth label. Structured human preference comparison, well-designed rubrics, and the honesty to say when a metric is a proxy for taste.
- Video data engineering at scale — ingest, transcode, storage, dedup and labelling — which is unglamorous and is a large fraction of the actual work.
- Working with artists and designers as peers. The people who find your model's failure modes fastest are the ones using it professionally, and engineers who treat that feedback as noise do not last.
Roles this kind of team hires
- AI Research Scientist
- Machine Learning Engineer
- AI Engineer
- MLOps Engineer
- Data Engineer (AI-focused)
- AI Product Manager
Linked titles have a full question bank with model answers. Unlinked ones are titles this employer uses that do not map onto one of our role guides.
How to prepare
Make something with the tools first
Generate a handful of clips, deliberately push until they break, and be able to describe the failure modes precisely — drift, identity collapse, physics that does not hold, motion that ignores the prompt.
Get concrete about generation economics
Know roughly what a few seconds of video costs in GPU time, and be able to argue for a specific tradeoff: fewer diffusion steps, lower preview resolution, a distilled model, or caching a shared prefix of the computation.
Design an evaluation for subjective quality
Explain how you would run preference comparisons that produce a usable signal: who judges, what they compare, how you control for prompt difficulty, and how you avoid optimising for a metric artists disagree with.
Read the research page properly
Their world-model framing is not decoration; it explains why the roadmap runs towards interaction and simulation rather than better clips. Know what they have published and how it connects to the shipped product.
Think about the asynchronous product surface
Their API submits tasks and polls for results. Be ready to design around that: queueing under load, progress feedback, failure and retry semantics, and what the user sees while a minute-long generation runs.
What to watch for
Common mistakes with this kind of employer
- Treating it as a text-LLM job. The modality is different in ways that matter — temporal coherence, compute per output and evaluation without ground truth — and language-model instincts transfer only partially.
- Ignoring cost. Proposing the highest-quality path for every generation, without a preview tier or any economics, reads as someone who has never paid a GPU bill.
- Dismissing the creative side. Candidates who talk about artists as end users to be managed, rather than the people whose judgement defines the quality bar, misread the company badly.
- Benchmark-only arguments. There is no leaderboard that settles whether a shot looks right, so an answer built entirely on published scores does not tell them what they need to know about you.
Sources
This guide is written from primary public material — Runway's own engineering writing, research pages, published job descriptions, and public talks. Nothing here is taken from a competitor's interview guide, and no specific claim about the loop is made without a source that supports it.
Last reviewed . Interview processes change without notice and vary by team, level, and location. Runway's own careers page is the authoritative description of how it hires; where this guide and that page disagree, that page is right.
Rehearse it before it counts.
Knowing what a company values is not the same as demonstrating it under time pressure. Book a 30-minute 1:1 and run the round you are least confident about.