# How to Prep for a World Labs Engineering Interview

Source: https://www.techinterview.org/post/3233477923/world-labs-interview-guide/
Updated: 2026-11-08 · techinterview.org

First, make sure you have the right World Labs: the one worth this prep is Fei-Fei Li's spatial-intelligence company in San Francisco, whose job posts name Justin Johnson and Ben Mildenhall as co-founders and whose product is Marble, a system that turns text, images, or video into navigable 3D worlds (worldlabs.ai, marble.worldlabs.ai). It does not publish an interview loop, and I found no first-person candidate reports, so treat any fixed sequence as a guess and confirm your stages with the recruiter. The open roles split three ways: research (generative modeling, 3D, SLAM, robot learning), platform and performance engineering, and graphics and rendering. Every role is in San Francisco and in person five days a week, and the posted base bands on the company's Ashby board run from about $200,000 to $350,000 with equity on top.

World Labs calls itself a spatial-intelligence company, which is a specific bet, not a slogan: its work is world models that perceive and generate 3D space instead of text. The concrete anchor is Marble, the product that turns a prompt or a photo into a 3D world you can move through, with uses the company's postings list across gaming, film, architecture, and robotics. Prep for the product and the research behind it, not a generic AI-lab loop.

## What World Labs is now, and why it is hiring

The company announced a $1 billion funding round on February 18, 2026, with AMD, Autodesk, NVIDIA, and Fidelity among the named investors; reporting put its total raised near $1.23 billion. World Labs did not state a valuation, and Bloomberg had reported in January 2026 that it was in talks around $5 billion, so treat that number as reported, not confirmed. Marble reached general availability on November 12, 2025: the launch post says the company is "making Marble, a first-in-class generative multimodal world model, generally available for anyone to use," after a run open only to early beta users. In January 2026 World Labs followed with the World API, a separate release that opened Marble to programmatic use. Shipping a 3D-generation product to real users takes people who can train the models, keep them fast on GPUs, and turn their output into something a renderer and a pipeline can use.

Spatial intelligence is the organizing idea to understand before you walk in. The claim is that a system that genuinely understands 3D space, geometry, physics, and how a scene changes as you move through it, is a different kind of model than one predicting the next token, and that real applications in robotics, design, and simulation need the former. Whether or not you buy the thesis, your interviewers do. A candidate who calls Marble "an image generator" rather than a world model will land wrong in the first ten minutes.

For peer calibration, the closest neighbors on this site are other world-model and generative-3D labs: [Odyssey](/post/3233477523/odyssey-interview-guide/), [Decart](/post/3233477462/decart-ai-interview-guide/), and [Luma AI](/companies/luma-ai-interview-guide/) hire for a similar mix of research and real-time 3D work, and the [AI-native company interview guides](/ai-startup-interview-guides/) hub collects the rest.

## The three tracks, and what they pay

World Labs runs a small board, and for an engineer the live openings fall into three families that interview differently. Sort out which one your req belongs to first; it decides what the interview probes and what the number on the offer looks like. The bands below are the posted ranges on the company's Ashby board, read in November 2026; Ashby shows each one as a structured salary range with a separate, unquantified "offers equity" flag instead of a written-out "US base salary range" sentence, so the figure is base only and the equity is a line you have to ask about.

| Role family | Representative open roles (San Francisco, in person) | What the loop leans on | Posted base band (World Labs Ashby board, read November 2026) |
| --- | --- | --- | --- |
| Research (world models) | Research Scientist, Generative Modeling; Research Engineer / Scientist for 3D, SLAM, and Robot Learning | Generative and 3D modeling, SLAM and geometry, robot learning, defending a research project in depth | $250,000–$350,000 base, plus equity |
| Platform and performance engineering | Performance Engineer (Inference, Training & GPU) | GPU performance, training and inference throughput, memory and latency, distributed training | $200,000–$300,000 base, plus equity |
| Graphics and rendering engineering | Pipeline Engineer (Graphics/3D); Rendering Systems Engineer (Simulation & Synthetic Data) | Real-time rendering, 3D asset pipelines, synthetic-data generation, engine integration | $200,000–$325,000 base, plus equity |

One note on older listings: search results still surface World Labs reqs on a Greenhouse board (titles like Real-time Graphics and Developer Infrastructure), but those URLs now return 404. The live board is Ashby, linked from the careers page on worldlabs.ai, and that is the only source I would trust for what is actually open and what it pays.

## What the loop probably looks like

World Labs publishes no interview loop, and I could not find a single first-person account on Glassdoor, Blind, or Reddit, so everything here is extrapolated from comparable world-model labs, not confirmed; get your actual stages from the recruiter on the first call. The sequence is likely unremarkable: a screen, a technical or research screen, an onsite of several rounds, and a fit conversation, run in San Francisco. What matters more is that at a lab this size the founders are often in the loop directly, so expect at least one conversation that is less about a rubric and more about whether you care about spatial intelligence.

Treat that shape as a frame, not a script. The one thing I would bet on is that the technical bar is set by the research, not by a generic coding screen: these roles ask you to reason about 3D geometry, generative models, SLAM, or GPU performance in depth, and a candidate who can only grind algorithm puzzles will stall on the second question. A [system-design](/category/system-design/) round here, if it appears, is more likely to be about a training or inference pipeline, or how synthetic data flows into a model, than about designing a web service.

## How to prep each track

On the research track (generative modeling, 3D, SLAM, robot learning), the hard part is a project you own. Be ready to defend one piece of work for an hour: the modeling choice, the evals, the failure cases, and what you would do with more compute. Know the papers this company is built on. Ben Mildenhall is the first author of NeRF, so neural radiance fields and 3D Gaussian Splatting are table stakes; for the SLAM roles, know a reference system like ORB-SLAM or DROID-SLAM and the geometry of estimating pose from real sensors. The stack is PyTorch and CUDA, so be fluent in both. Come in with opinions about what world models still get wrong.

On the platform and performance track, the weight lands on making large models train and serve fast: GPU kernels and memory, throughput bottlenecks, mixed precision, multi-GPU training, and inference latency tight enough for an interactive product. Expect CUDA, Triton, FlashAttention-style kernels, and a distributed-training stack like FSDP or DeepSpeed to come up by name, and have a concrete story about a bottleneck you found and the number you moved. On the graphics and rendering track, expect real-time rendering, 3D asset pipelines, and the plumbing between a generated world and an engine that has to display it; synthetic-data generation sits here too. Co-founder Christoph Lassner, who writes for the company on real-time graphics and comes out of the PyTorch3D and neural-rendering world, is a fair read on the bar: know real-time 3D Gaussian Splatting, differentiable rendering, and a graphics stack such as OpenGL or Vulkan alongside PyTorch3D. Across all three a coding screen is still plausible, and the usual [coding patterns](/algorithm-patterns-cheat-sheet/) cover it, but it will not be the part that decides the loop.

No real question bank is public, so the prompts below are the kind the roles imply, not leaked questions. Rehearse them out loud until the answers are specific:

- Walk me through how a NeRF or 3D Gaussian splatting scene is represented, and where each representation breaks down.

- Given a single video, how would you estimate camera pose and build a consistent 3D map, and how do you keep drift from accumulating?

- A generated world looks right from the training viewpoint but falls apart from a new angle. How do you diagnose that?

- You need inference latency low enough for the product to feel interactive. Where do you look first, and what do you trade away to get there?

- How would you generate synthetic data to teach a model to reason about a space it has never seen?

## What the roles pay, and how to verify it

Because World Labs is in San Francisco, its postings carry California-mandated pay, which is why the table quotes real bands instead of guesses. Research roles post the highest, roughly $250,000 to $350,000 base; the performance-engineering role sits near $200,000 to $300,000; graphics and rendering roles land around $200,000 to $325,000. Every one adds equity, but Ashby marks it only as a flag with no percentage, so the grant is the number you have to get from the recruiter. At a company that just raised a billion dollars, that grant is the part most worth negotiating.

Pin the real number from more than one angle. The posted base band for your exact req is the anchor; Levels.fyi by level is a secondary read, and for a company this new expect few or no verified entries, so treat it as a rough signal, not a figure. Get the equity grant, the strike price, and the last round's preferred price from the recruiter early, model the package with the [total comp calculator](/total-comp-calculator/), and anchor any [negotiation](/post/3233474669/salary-negotiation-2026/) on your level and the posted band instead of the funding headline.

Two things the board hides, both of which matter to a senior candidate. The postings attach no level to the bands, so the $350,000 research ceiling most likely tops out a senior IC, not a staff or principal tier the board never posts; ask where your level sits and whether anything exists above it. And since equity is only a flag with no number, make it comparable by asking the grant's dollar value at the last round's preferred price, because a raw percentage means nothing until you know its strike.

## One thing to settle before you sign

The in-person policy is not a footnote. Every open role says five days a week in the San Francisco office, which is a real filter: if you need remote or hybrid, raise it on the first call instead of hoping it bends later, because a small research lab that wants everyone in the room tends to mean it. The postings I read are also silent on visa sponsorship and relocation; if either matters, ask the recruiter directly. Ask too what "research" versus "engineering" means for the specific team you would join, since the titles blur here and a Research Engineer role can be mostly infrastructure or mostly modeling depending on the group. For the behavioral parts, keep your stories tight with the [STAR structure](/post/3233460379/behavioral-interview-questions-2026-star-method-amazon-leadership-principles-and-winning-answers/), and use the [study plan generator](/study-plan/) to split the week between research depth and the coding screen. The [AI-startup interview difficulty index](/ai-startup-interview-difficulty-index/) sets expectations for a lab at this stage: small, research-heavy, and far more interested in whether you think in three dimensions than in how fast you can reverse a linked list.