Prepping for an AI engineering interview usually looks like this: twelve tabs of Glassdoor, a Reddit thread from 2023, three YouTube videos, a half-read paper, and no idea what this specific company will actually ask.
So I went through 270 public interview reports, the write-ups and videos where candidates name the company and the question, and sorted them by company. The first thing that jumps out: every top lab has its own personality. Prepare for "AI interviews" in general and you'll be ready for none of them.
Here's what each one asks, with the real questions.
OpenAI: design it for real constraints
9 reported questions · 6 senior, 3 mid
OpenAI's list is mostly system design with hard constraints attached: latency, cost, permissions, tenants.
- "Design a RAG system that has to meet strict latency and cost targets while serving many tenants."
- "Design an agent for searching internal documents that respects permissions, cites sources and abstains when it should." (also reported at Google and Lyft)
- "Design a harness for running a tool-using agent safely, and an evaluation platform on top of it."
- "Here's a small transformer with a few bugs. Fix them, then add a KV cache for decoding."
What they're testing: can you build something that holds up in production, not just in a demo? "Abstains when it should" and "evaluation platform on top" are the giveaways: they want guardrails and measurement designed in from the start.
Anthropic: do the maths, out loud
10 reported questions · 7 senior, 3 mid
Anthropic's list reads like an infrastructure exam. Expect back-of-envelope estimates on the spot.
- "You need to copy 500 GB of weights to 1,000 GPU workers over shared 10 Gbps links. What's the lower bound on time, and what would you build?"
- "Estimate an FFN layer's FLOPs, parameter and activation bytes, and the communication from sharding it. Which limit binds: compute, HBM or network?"
- "Debug a GRPO training loop. What is the probability ratio, and what goes wrong with it?"
- "Turn an LLM into a binary classifier using its log-probs and log-sum-exp."
Here's what a strong answer to the first one looks like:
Each worker has to receive 500 GB through its own 10 Gbps link, so 400 seconds is the floor. Serve everyone from one machine and it takes about 4.6 days. The trick is to make every receiver also upload: chunk the weights and pipeline them down trees, or swarm them BitTorrent-style, and you get close to 400 s. Then cover topology (one copy per rack uplink), checksums and resuming after failures.
What they're testing: whether you think in bytes, bandwidth and bottlenecks before you think in boxes and arrows.
Mistral AI: build it from scratch
16 reported questions, the most of any lab · 10 mid, 5 senior, 1 junior
Mistral goes deep on model internals. You will write code.
- "Given logits from a toy language model, implement greedy, top-k and top-p decoding in NumPy."
- "Implement batched multi-head self-attention with a causal mask."
- "How does RMSNorm differ from LayerNorm, and why do recent models use it?"
- "Explain FSDP and ZeRO stages 1, 2 and 3. How do you overlap computation with communication?"
- "Your RAG chatbot keeps giving wrong answers. How do you find out why, and fix it?"
What they're testing: whether you know what's inside the model, down to the norms, the KV cache and the optimiser. If you've only ever called an API, this is where it shows.
Meta: agents, end to end
9 reported questions · 7 senior, 2 mid
Meta's senior questions are big, open design problems, mostly about agents and coding assistants.
- "Design a code-review agent that combines LLM orchestration with deterministic tools."
- "How do you keep a coding agent from leaking private code, both through training and at inference time?"
- "Design an LLM coding assistant end to end: data, model and evaluation."
- "Implement self-attention, give its time and memory complexity, and describe the efficient variants."
What they're testing: whether you can scope a large system, know where the LLM ends and the deterministic code begins, and treat security as part of the design.
Google: post-training and safe agents
11 reported questions · 6 senior, 4 mid, 1 junior
- "How do you detect and prevent reward hacking during LLM post-training?"
- "Design LLM summaries of e-commerce reviews that hold up against prompt injection and don't leak personal data."
- "Design a scalable service where an AI agent plans work and calls approved tools without ever bypassing deterministic controls."
The one question to prepare for everywhere
"Implement batched multi-head self-attention with a causal mask."
It was reported at Mistral AI, ByteDance, Meta, Apple and Uber, more companies than any other question in the set. Add the KV-cache variants that turn up at OpenAI, Netflix and Mistral, and writing attention from a blank file is the highest-return hour of prep you can do.
How a prep session works on LLM Dojo
I put all of this into LLM Dojo, so preparing for a specific company is one tab instead of twenty. Here's what a session looks like:
1. Open your company's page. Every question candidates reported there, split by level, with a link to each original report.
2. Read the question and answer it out loud. One or two minutes, like the real thing. Out loud is where the gaps show.
3. Open the model answer and compare. Every answer is short enough to say in an interview and lists its sources: papers, books, measurements.
4. Didn't click? Open the lesson. Every topic links to a lesson with interactive labs, so you can see the idea working instead of memorising it: race prefill against decode, size a KV cache with sliders, step through BPE, run GPT-2 token by token in your browser.
5. Let it come back. Sign in (free) and the checks you miss return after 1 day, then 3, then 7, until they stick.
Everything in one place
- 492 interview questions, junior to senior, each with a model answer
- 262 of them tagged to the company where candidates reported them, across 58 companies: OpenAI, Anthropic, Mistral, Meta, Google, Microsoft, Amazon, Apple, NVIDIA, Cohere, Scale AI, Databricks, Perplexity and more
- 7 modules, 41 lessons covering everything those questions test: model internals, inference, prompting, evals, RAG, agents and production
- 70+ interactive labs and a neural network playground
- Spaced review, so what you learn this week is still there on interview day
- Free. The interview hub needs no account
Got an interview coming up at a company that isn't covered, or were you asked something that isn't in the bank? Drop it in the comments and I'll add it, with a model answer.
