Career & Practice 🗂️

A Public Study Stack for Frontier AI Research Roles

Four people who actually went through frontier-lab research interviews — at OpenAI, DeepMind, and elsewhere — published what they studied. A Stanford course put its curriculum online with the assignments included. None of it was written as a unified guide, but read together it forms something close to a real preparation trail for Research Scientist, Research Engineer, Applied Scientist, and technical MTS roles.

FrontierAGI Team

September 24, 2026 · 16 min read

Where this comes from: this piece draws on and synthesizes a September 22, 2026 write-up by Shilpa Mitra on the WebAfterAI newsletter, which itself surveys four public first-person accounts plus a Stanford course. All source links below point to the original authors' own pages, not to any secondary summary — that's deliberate, since the whole point of this piece is that the primary material is public and worth reading directly.

Why This Exists Alisa Liu — OpenAI Stanford CS336 Silvia Sapora — DeepMind Yong Zheng-Xin Nathan Lambert, 2022 Public Interview Repos Putting Together a Stack Full Reference List

Why This Exists

Interview preparation for frontier AI research roles is usually invisible — people do it privately, get the offer or don't, and the specifics disappear. What makes this particular moment interesting is that several researchers who went through the process recently chose to publish the actual material: not polished "how I got the job" threads, but the notes, checklists, and topic lists they built while preparing. Read individually, each is one person's account. Read together, they start to look like a real curriculum for what frontier labs currently expect a research hire to know.

Alisa Liu: 57 Interviews, Then the Notes

Alisa Liu joined OpenAI's pretraining team after finishing an NLP PhD at the University of Washington, having worked on tokenization, data generation, and inference-time algorithms. Getting there took 57 interviews across 11 companies, plus 46 recruiter calls and 16 post-offer conversations. Rather than a summary thread, she published her full job-search write-up, along with the LLM notes and math notes she actually built while studying.

The interviews were broader than her research. Despite six years of focused NLP research, her rounds spanned ML coding (the most common type, by her account), general coding, technical discussions, research discussions, behavioral interviews, math, and job talks — implementing an architecture in PyTorch in one round, reasoning through a backward pass in NumPy in another, jumping between positional encodings, parallelism, PPO, and GRPO in a third. Her preparation ended up resembling a return to coursework more than answer-polishing: continuously building notes, drawing diagrams, and practicing implementation from scratch.

Her notes are useful less as a document to memorize and more as a checklist to test yourself against. Take attention as an example: can you explain why it works, write self-attention without looking anything up, implement causal masking, derive the dimensions, explain cross-attention, discuss the memory and compute costs, and explain what FlashAttention is actually doing? The gap between recognizing a concept and being able to derive, implement, debug, and discuss it is where a meaningful share of this preparation lives.

Where Liu Started: Stanford CS336

The first major resource in Liu's own preparation was Stanford CS336: Language Modeling from Scratch — not a course about calling an LLM API, but one that walks through the actual stack: tokenization, transformers, optimizers, training, GPUs, kernels, distributed training, scaling laws, inference, data, evaluation, alignment, and reinforcement learning. The course's assignments are public, and the first assignment has you implement a tokenizer, transformer architecture, optimizer, and training loop needed to train a small language model — with Stanford's own materials explicitly recommending AI autocomplete be turned off while doing it.

Why that recommendation matters beyond the course itself: implementing and debugging transformers comes up often enough in interviews that it needs to become muscle memory, not something you can reconstruct with an assistant's help in the room. Discovering in a 45-minute coding round that your PyTorch fluency was actually your editor's autocomplete fluency is exactly the failure mode this is meant to prevent.

Silvia Sapora: The DeepMind Version

Silvia Sapora interviewed for Research Scientist roles at DeepMind, Isomorphic Labs, Cohere, Meta, and several other companies before joining DeepMind, and published ML Job Interviews: The Ultimate Guide afterward. Where Liu's write-up gives you a preparation process, Sapora's gives you a large, specific technical topic list: RL concepts (PPO, GRPO, GAE, DPO, policy gradients, model-based RL), LLM concepts (FlashAttention, MoE, scaling laws, RoPE, tokenization, RLHF), distributed-training concepts (tensor parallelism, FSDP, pipeline parallelism), plus classical ML, probability, statistics, and linear algebra.

Implement a transformer end to end.

Causal, self, and cross attention — including FlashAttention and the attention backward pass.

An MLP forward and backward pass, and a simple training loop, from scratch.

Tensor reasoning practice via Tensor Puzzles, and systems-side depth via the JAX Scaling Book.

That's a coding baseline — a specific bar rather than the vague instruction to "learn deep learning."

Yong Zheng-Xin: What Didn't Fit the Pattern

Yong Zheng-Xin, a Brown PhD student moving toward AI safety research, published a write-up that's useful precisely because it doesn't repeat Liu's or Sapora's advice. His core observation: published papers may matter mostly for getting into the interview pipeline in the first place — once inside it, candidates can be evaluated on very different skills, including system design, parallel programming, how well they use AI agents as tools, and paid work trials lasting days. He also makes an encouraging point for anyone changing research direction: having fewer papers in a new area may matter less than being able to demonstrate real understanding of the problems that area's team actually cares about.

The warning label this account provides: don't optimize an entire preparation plan around one imagined interview template, whether that template comes from Liu's OpenAI experience or Sapora's DeepMind one. Wildcard rounds — system design, agent-usage evaluation, multi-day work trials — are common enough that a preparation plan built only around "the" frontier-lab interview format will have real gaps.

Nathan Lambert: The Earlier Version of the Same Journey

Before any of the above, there was Nathan Lambert's 2022 write-up of his own job search while leaving Berkeley AI Research for a top industry Research Scientist role — covering where interviews actually came from, how networking shaped opportunities, and how different research organizations behaved during the process. Liu has said she used Lambert's timeline as direct inspiration for documenting her own search, which makes this less a fifth unrelated data point and more the first link in the chain that produced the other four.

Read together, these four accounts stop the hiring process from looking deterministic. There's no clean sequence where a PhD plus a few papers automatically becomes an offer. A closer approximation: strong research, getting into the pipeline, coding and ML breadth, research judgment, math and systems depth, wildcard rounds, team match, timing, headcount, and negotiation, all layered on top of each other. Research credentials open doors; they don't replace preparation.

Public Interview Repositories Worth Keeping Around

AI/ML Interviews — built from its author's own preparation after offers from Meta, Google, Amazon, Apple, and Roku; covers general coding, ML coding, ML fundamentals, LLMs, ML/GenAI system design, agentic systems, and behavioral interviews.

Deep Learning Interviews — hundreds of solved questions across probability, information theory, optimization, neural networks, and deep learning, for rapid technical drilling.

ML Interview — a smaller collection covering statistics, linear algebra, deep learning, large-scale ML, and system design, built while preparing for ML Engineer, Applied Scientist, and Quant Researcher roles.

These are better used as question generators than as material to read front to back — if a topic makes you uncomfortable when you hit it, that's usually the one worth actually studying.

Putting Together a Stack

One reasonable way to sequence this material — not the only way, and not a substitute for reading all four accounts in full.

A rough sequence: start with CS336 to understand the whole stack; use Liu's notes as a checklist to find gaps in breadth; implement the core pieces — transformer, attention, training loop — from scratch rather than reading about them; use Sapora's topic list to stress-test what's still shaky; use the interview repos for rapid-fire practice once the fundamentals hold; and read Liu, Sapora, Zheng-Xin, and Lambert together to understand the actual hiring process, not just the technical bar. Then, periodically, turn AI assistance off and see what's actually left. Liu used ChatGPT and Claude heavily while learning, but specifically recommends practicing coding without AI assistance before interviews — it's easy to underestimate how much of your fluency has quietly become the tool's fluency instead of your own.

🎯 The Bottom Line

You don't have to be interviewing next month for this material to be worth keeping. Liu has said directly that interview preparation made her better at research — studying broadly exposed gaps she'd learned to route around during a PhD, and filling them widened the set of problems she felt comfortable thinking about. A researcher can spend years going deep in one narrow area while getting rusty on linear algebra, systems, optimization, or the parts of modern LLM training outside their own work. These four accounts, read together with CS336, amount to something close to a frontier-AI technical curriculum built backwards from what strong labs currently expect a researcher to know — imperfect, since no single interview process is universal, but public, current, and written by people who actually went through it. The highest-value move probably isn't reading their notes passively. It's taking their topic lists, finding what you can't explain or implement comfortably, and building your own version until you can.

🔗 Full Reference List

Shilpa Mitra, "A Public Study Stack for Frontier AI Research Roles" — WebAfterAI, Sep 22, 2026 Alisa Liu — job-search write-up Alisa Liu — Book of LLMs Alisa Liu — math notes Stanford CS336: Language Modeling from Scratch CS336 — Assignment 1 (public) Silvia Sapora — ML Job Interviews: The Ultimate Guide Tensor Puzzles JAX Scaling Book Yong Zheng-Xin — job-search write-up Nathan Lambert — AI PhD job hunt (2022) AI/ML Interviews (repo) Deep Learning Interviews (repo) ML Interview (repo)