Thinking Machines Lab runs a headcount you could fit in a mid-size startup’s engineering org while sitting on a $12 billion valuation and a $2 billion seed round that Andreessen Horowitz led in July 2025. Nvidia, AMD, Cisco, and Jane Street are on the cap table. That math sets the tone for hiring: with a team in the fifty-to-sixty range and no product revenue to speak of yet, a single engineer is a real fraction of the whole company, and the interview reflects it.
Mira Murati started the lab in February 2025 after leaving her job as OpenAI’s CTO, and pulled in around thirty researchers and engineers from OpenAI, Meta AI, and Mistral before the doors were really open. John Schulman, an OpenAI cofounder and the person behind much of the early RLHF and PPO work, is chief scientist. The founding bench read like a roll call of people who shipped ChatGPT, PyTorch, Fairseq, OpenAI Gym, and Segment Anything. If you interview here, some of the people across the table wrote the tools you trained on.
Two things happened in the year after launch that a candidate should understand before walking in. The lab shipped Tinker in October 2025, an API for fine-tuning open-weight models that leans on LoRA so small teams can post-train without owning a GPU cluster. Then in May 2026 Murati showed “interaction models,” a bet on full-duplex systems that take audio, video, and text as one continuous stream instead of turn by turn. Through all of it the talent war got ugly. Meta hired away five people connected to the founding group in mid-2026, and one reported offer for Andrew Tulloch carried a headline figure around $1.5 billion over several years, while Barret Zoph and Luke Metz went back to OpenAI in January 2026. A planned raise near a $50 billion valuation fell apart around the same time.
That matters for your interview because the lab is hiring to replace deep, specific talent, and it screens for the same thing it lost: people who are unusually strong at one hard problem rather than broadly fine at everything.
Member of technical staff, and the research-engineer split
Almost everyone technical carries the same flat title, member of technical staff, which is common at frontier labs and drops the ladder-climbing signals you would read at a big company. Under that title the real division is between research scientists, who own the direction of a research bet and are expected to publish and set the agenda, and research engineers, who make large models train fast and serve reliably and who often drive the experiments that make a paper real. The line is blurry on purpose. The public research-engineer description covers pre-training methodology, large-scale experiment design, and inference and deployment performance in the same breath.
What that means in practice: a research engineer here is expected to read a paper and reproduce it from scratch, config and all, rather than run a setup someone else wrote. A research scientist is expected to write the training loop when it matters, not hand it off. The interview probes both sides no matter which door you came in through, weighted toward whatever you are deepest in.
The loop, round by round
The lab does not publish its process, and it is small enough that the loop flexes per candidate and per team. What follows is the realistic shape, drawn from how frontier research labs this size run and what the roles ask for. Expect the whole thing to move faster than a big-company loop, sometimes a week or two from first call to offer, because the founders sit close to every decision.
| Round | Format | What it screens for |
|---|---|---|
| Recruiter or founder screen | 30–45 minute call | Depth and ownership of past work; whether your spike matches an open need |
| Technical screen | Live coding, about 60 minutes | Practical ML coding: attention, data loaders, debugging a broken training loop |
| ML and research depth | 60–90 minute discussion | First-principles grasp of your own work, RL, optimization, and scaling tradeoffs |
| Systems and infra (research-engineer track) | About 60 minutes | Distributed training, GPU memory math, parallelism, and inference performance |
| Project or paid work trial | Take-home or on-site day | A self-contained problem that mirrors day-to-day research engineering |
| Founder and fit conversations | Several short chats | Operating inside a tiny, high-context, fast-changing team |
The recruiter or founder screen is short and unusually technical for a first call. Someone will ask what you actually built, and they will know enough to tell whether you led it or stood next to it. Have a real answer to “what is the hardest bug you have chased in a training run,” and be ready to name the one area you are deepest in.
The technical screen is a coding round, but not a LeetCode round. The tasks look like the job: implement multi-head attention without reaching for a framework’s autograd, write a data loader that shuffles a dataset too large to fit in memory, or take a training loop that silently emits NaNs and find where the numerics go wrong. Clean, correct code under mild time pressure counts for more than a clever trick.
The research and ML-depth round is where offers are won and lost. It usually opens with your own work: pick a project, walk through it, then keep answering as the interviewer pushes on the parts that did not work. They are testing whether you understand your results or just remember them. Expect follow-ups that go to first principles:
- Derive the policy-gradient estimator, then explain why PPO clips the probability ratio.
- Why does LoRA recover most of full fine-tuning quality, and where does it break down?
- Your eval loss is flat while training loss keeps dropping. Walk me through what you check.
- On a fixed compute budget, would you train longer or scale the model, and how do you decide?
Research-engineer candidates get a systems round built around making large models go fast. The questions are concrete: your all-reduce is the bottleneck at 512 GPUs, what do you look at first; estimate the KV-cache memory for a 70-billion-parameter model at an 8k context window; a model that will not fit on one node, how do you split it and what does each strategy cost you in communication. Back-of-envelope fluency with GPU memory, bandwidth, and parallelism strategies is the thing being measured, not memorized trivia.
Some candidates get a project or a paid work trial instead of, or on top of, a whiteboard onsite. A short, self-contained problem that mirrors real work tells the team more than a day of interviews, and it tells you whether you would actually want to do the work. The final conversations are with founders and cover fit in the specific sense of whether you will operate well in a small, high-context, still-forming company where the roadmap changed twice this quarter.
What the offer looks like
Comp at Thinking Machines is private and the equity is illiquid, so treat any number you hear with suspicion. Member-of-technical-staff packages at frontier labs have been reported across a wide band, often quoted from the mid-six figures into seven figures of total comp, weighted heavily toward equity rather than base. The Meta poaching numbers that made headlines, including the reported figure for Andrew Tulloch, were retention-grade offers for founding-level people and tell you nothing about a normal individual-contributor package. Put plainly: base salaries here sit competitive with other top labs, and the real question is what you believe the equity is worth in a company that just watched a $50 billion round evaporate. Check levels.fyi and recent reporting for current ranges rather than trusting a recruiter’s midpoint, and price the equity as the lottery ticket it is.
How to actually prepare
The candidates who do well here are not the ones who ground through a problem set. They are the ones who can go three questions deep on something real. Pick the area you are strongest in, whether that is RL, distributed training, inference optimization, or evals, and get to where you can derive the core results from scratch and argue the tradeoffs like you have been burned by them. Reproduce a paper end to end; the gap between reading it and making the numbers match is exactly what the ML-depth round probes.
Read what the lab has put out. Tinker’s design choices around LoRA and fine-tuning APIs, and the interaction-models framing from May 2026, are fair game for “why do you think we built it this way” questions, and having an opinion beats having read the launch post. If you are on the engineering track, spend real time with the arithmetic of large-model training and serving: FLOPs, memory, bandwidth, the cost of each parallelism scheme. Those numbers come up because the job is made of them.
What the lab is buying is a person who raises the technical ceiling of a room that is already very high. Walk in with one thing you know better than almost anyone, and be ready to be wrong out loud about everything else.
