company guides

Periodic Labs interview guide for ML and research roles

If you’re interviewing at Periodic Labs, start with one fact that shapes everything else about the loop: the company is trying to build an AI scientist and the robotic lab that scientist runs experiments in. Its models propose new materials, a powder-synthesis lab in Menlo Park makes and measures them, and the results feed back into training. That founding team is credited with work behind ChatGPT, DeepMind’s GNoME materials model, and OpenAI’s Operator agent, per the company’s own site and SiliconAngle’s reporting on the seed round. Co-founders Liam Fedus (a former OpenAI VP of research) and Ekin Dogus Cubuk (who led materials and chemistry work at DeepMind) set the bar for who they want to hire: people who can do frontier ML and stay grounded in what a physical experiment actually shows.

Because the company is young and keeps a low public footprint, there’s no Glassdoor or Blind thread describing its loop. Everything here about round structure is labeled as inference. The company facts, role titles, and pay bands are drawn from primary sources and attributed. For reference points from peers whose loops are better documented, the AI-native company interview guides hub and the AI-startup interview difficulty index are a good place to start.

What Periodic Labs builds, and why it shapes the questions

The product is a loop, not a single model. Periodic Labs describes itself as building models and autonomous labs to accelerate science, with an early target of discovering superconductors that work at higher temperatures than today’s materials. The data flywheel matters for interviews: the company bets that experimental data generated in its own lab beats models trained only on public web text. So a strong candidate is someone who can train and scale models and also reason about messy, real measurements coming off X-ray diffraction and thin-film equipment.

That dual nature is why the hiring split is unusual. Most open roles sit on the science side (experimentalists, computational scientists, process engineers), with a smaller frontier-model and infrastructure group. If you come from a pure LLM background, expect at least one conversation about how your work connects to a physical experiment you can’t hand-wave past. This is closer to Physical Intelligence or Sakana AI than to a chat-app startup, in that the science is the point rather than a wrapper.

The engineering and research roles open right now

As of November 2026, the company’s Ashby board lists around 30 open roles (excluding a catch-all “don’t see your role” posting), the large majority based in Menlo Park, with a few also open to Montreal. The engineering and research titles most relevant if you’re reading this: ML Systems Engineer, Software Engineer, Product Engineer, Research Scientist for Scaling RL, Research Scientist/Research Engineer for Midtraining, Research Scientist for Data, Research Engineer for Semiconductor and for Lab Automation, and several Computational Scientist roles.

The job descriptions name the stack directly, which tells you what to prepare. The ML systems group references work behind Megatron-LM, SGLang, Liger Kernel, TorchRec, CleanRL, TorchRL, and JAX-MD, and asks for large-scale training, inference, and reinforcement-learning infrastructure plus GPU-kernel and communication optimization. The RL research roles reference RL-scaling work such as ScaleRL and runs like the company’s own Periodic Neon model, and they ask how reinforcement learning scales with compute, model size, and reward quality. Midtraining roles are about curating scientific data, building evals that track downstream task performance, and distillation. If you’re targeting the systems side, the same muscles show up in guides for ML infrastructure shops like Modal, and a solid grounding in system design helps for the distributed-runtime questions.

The Periodic Labs interview loop, inferred not reported

No candidate has published a Periodic Labs loop, so treat the table below as a working model built from the open-role requirements and from how teams with this OpenAI and DeepMind pedigree usually interview. Expect it to shift by role and to move fast, because the company says plainly that it operates at the pace the frontier demands.

Periodic Labs interview stage (inferred, Nov 2026) What it most likely covers Source / status
Recruiter screen Motivation for AI-for-science specifically, role fit, logistics, and comp expectations Inferred from open roles and frontier-lab norms
Technical phone screen For engineers, practical coding plus ML systems reasoning; for research roles, a close walkthrough of a paper or project Inferred from open roles and frontier-lab norms
Main technical onsite Distributed training or RL infrastructure, GPU and throughput tradeoffs, or an RL-scaling or evals discussion depending on team Inferred from job-description requirements
Science-grounding round How your work connects to real experiments and noisy measurement data, likely with a computational or experimental scientist Inferred from the model-plus-lab product
Values and founder round Ownership, speed, and why this mission; likely time with a founder given company size Inferred from team size and stated culture

Two things to keep in mind. These five stages are a likely shape, not a confirmed agenda, and at a company moving this fast they may be compressed into fewer, sharper conversations. The science-grounding conversation is the part the role requirements suggest candidates would underprepare, because it doesn’t look like a standard ML interview and you can’t bluff your way through a question about what a diffraction pattern is actually telling you.

What Periodic Labs screens for

Reading across the engineering and research descriptions, a few signals repeat. Depth in large-scale training and RL that goes past using a framework into reasoning about policy optimization, credit assignment for long-horizon tasks, and training-versus-inference mismatch. Systems performance: memory, kernels, and communication across thousands of GPUs, with correctness that holds as workloads scale. Data and evaluation judgment: the ability to build benchmarks that correlate with real scientific outcomes, working next to physicists and chemists rather than around them. And ownership at startup pace, which in practice means you’re expected to take a vague problem to a working result without much process.

Expect the behavioral bar to stay real even if the loop is short. Be ready to tell a tight story about a time you owned an ambiguous system end to end. If you want structure for that, the STAR-method behavioral guide is a good scaffold, and the broader shift toward interviewing in the AI era applies doubly at a company whose entire thesis is autonomous research.

Practice prompts drawn from the job descriptions

These are not reported interview questions. They are practice prompts I built from the public role requirements, phrased the way the work is described, so you can rehearse the right muscles:

  • Walk through how you would scale an RL training run from a single node to thousands of GPUs, and where correctness tends to break first.
  • Design an evaluation that actually predicts whether a model got better at a scientific task rather than at a convenient proxy benchmark.
  • You get noisy experimental results that disagree with the model’s prediction. How do you decide whether the model, the data, or the experiment is wrong?
  • Optimize a serving path for a large model under a tight latency budget; name the kernel and communication tradeoffs you would reach for.
  • Build a sandbox where an AI agent can run and validate experiments without corrupting shared lab state.

For the frontier-model research roles, read up on ScaleRL and Periodic Neon so you can speak to both, since the job descriptions point at that kind of RL-scaling and post-training work directly. Peer frontier labs interview similarly; the Mistral guide gives a feel for how research-heavy loops probe depth rather than trivia.

Compensation: what is actually posted

Thanks to California pay-transparency rules, the Periodic Labs job board carries real salary bands, so these numbers are posted by the company rather than guessed, and you can check them per role yourself. As of November 2026, the ML Systems Engineer and Product Engineer roles post about $250k to $350k base plus equity in Menlo Park, and the Software Engineer role posts the same $250k to $350k plus equity. The Scaling RL research role posts roughly $225k to $350k base plus equity, open to Menlo Park or Montreal. Midtraining and Data research roles land around $250k to $350k plus equity. More hardware-adjacent roles sit lower: Research Engineer for Semiconductor around $220k to $300k, Lab Automation around $200k to $250k, and a Thin Films postdoc around $180k to $225k, all plus equity. Research and engineering roles generally offer visa sponsorship, and a few (Scaling RL, Data) are open to Montreal as well as Menlo Park, which matters if you need Canadian work authorization; the hourly lab-technician roles offer no sponsorship.

The headline cash is competitive but not eye-watering by frontier-lab standards, which tells you the real upside is equity in a company that reportedly went from a seed valuation near $1.3B to much higher talks within a year. Treat the equity as the number that matters, and get specific before you sign. A few questions worth asking at this stage:

  • The liquidation preference stack: how much gets paid out ahead of common stock before you see anything in an exit.
  • The current strike price and the date of the latest 409A valuation.
  • The vesting schedule, any cliff, and whether equity refreshers are on the table.

Model the whole package with a total-comp calculator and read the salary-negotiation guide before you discuss a band, because at a startup this early the equity terms and preference stack move your real outcome far more than base.

One more sourcing note so you can calibrate. The $300M seed closed in September 2025 at roughly $1B pre-money, the figure Bloomberg expected per SiliconAngle, which works out to about $1.3B post-money once you add the raise. The larger 2026 numbers are talks, not closes: Bloomberg reported discussions near a $7B valuation in March 2026, and Forbes reported advanced talks for at least $500M at about $7.5B in May 2026. As of November 2026 there was no announcement that either had closed, so weigh any recruiter’s framing of the company’s valuation against that uncertainty rather than taking a single headline as settled.

If this is your shortlist, the cleanest prep is narrow and deep: pick the one team you’re targeting, learn its named tools cold, and have a real story about connecting a model to a physical measurement. That last part, connecting a model to a physical measurement, is where the role requirements suggest most strong ML candidates would come up short.

newsletter

What's actually being asked right now

Interview patterns & comp trends, straight to your inbox.

No spam. Unsubscribe anytime.

newsletter

What's actually being asked right now

Interview patterns & comp trends, straight to your inbox.

No spam. Unsubscribe anytime.

1972 Soviet postage stamp commemorating the Mars 2 probe

worth a read

Mars For The Rest of Us — a weekly-or-more deep dive on the technical side of Mars exploration: rocket propulsion, microbiology, mission architecture, and everything in between. Written by Maciej Ceglowski.

Read it on Substack →
Scroll to Top