Google DeepMind Interview Notes

Practice problems modelled on Google DeepMind's coding, ML-knowledge, ML-design, research and behavioural interviews, plus a guide to the hiring process.
Full problem statements, worked solutions, and code that CI runs on every push.

English · 中文

Note

Unofficial. This project is not affiliated with, endorsed by or sponsored by Google DeepMind or Google. The problems are reconstructed from second-hand accounts of interviews, mostly from roughly the past year, plus a few older question types that still recur, and written up from scratch: statements, examples, solutions and code are all original. Far fewer detailed accounts of Google DeepMind interviews are public than for some other companies, and many give only the topic of a question ("a hard BFS variant", "the maths of a core LLM concept"); such pages are original problems built on that topic. Treat every page as practice on the kind of problem you may meet, not as the exact question. If you believe something here should not be public, open an issue and it will be taken down.

What is in it

Section	Pages	What they cover
Hiring process guide	1	The two ways into Google DeepMind, every stage from the recruiter call to the offer, how the old quiz changed, timelines, posted salary ranges, the Student Researcher Program, and the rules on AI tools during interviews
Coding	17	Graph search in several forms (a depth-first warm-up on an edge list, state-space BFS with path counting, games on graphs, a unit-conversion graph), Python generators with unit tests and weighted sampling, streaming frequency counting, binary search on bitonic arrays and monotone functions, bit packing, snapshots, k-d trees, a code-review round, and ML implementation: attention with a KV cache, backpropagation, EM, focal loss, debugging a training loop
Quiz	7	The oral knowledge round: ML fundamentals, LLM fundamentals, reinforcement learning for language models, two maths pages (probability, statistics and linear algebra; matrix computation, statistical tests and information theory), computer-science fundamentals, and explaining what a piece of code computes
System design	8	The Gemini app, an evaluation system for an LLM assistant, predicting a reaction factor for pairs of molecules, a retrieval-augmented agent, a recommendation feed, training a model that does not fit on one accelerator, a robot-learning data flywheel, and predicting which data-centre machines need replacing
Behavioral	5	Recruiter screen, research deep dive, research talk and paper defence, Googleyness and culture, team-lead and hiring-manager interviews, and the product-manager loop

What sets the pages apart:

Complete statements. Each problem is written out in full, with definitions, function signatures and worked examples, so you can attempt it without guessing what was meant. Most problems come in three parts that build on each other, the way the round does.
Solutions you can check. Every coding, quiz and system-design solution ends with a collapsed block of runnable checks: asserts on the examples, a brute force built independently from the statement wherever one can be written, numerical verification of every derivation, and, for system design, the estimates recomputed in code. CI runs the code of all pages on every push.
Pitfalls where they happen. Instead of a separate list of mistakes, # NOTE: comments sit on the line the mistake would be made.
The oral knowledge round, written down. The quiz pages give each question the way it is asked and the answer the way you would say it, followed by the derivation and a check in code.
The process, not only the questions. The hiring process guide separates what Google DeepMind publishes (linked) from what candidates report, stage by stage.
Study tracks. The roadmap orders the pages for research (RS / RE), software engineering, applied AI and ML engineering, student researchers and interns, and product managers.
English and Chinese. Every page exists in both languages with identical code.

How to use it

1. Read the process guide

Start with the hiring process guide: which door you are entering by (Google DeepMind's own posting, or Google's general loop followed by team match), which rounds that means, and how long it takes. "Why DeepMind" comes up in nearly every round, so the behavioural pages are worth starting early.

2. Pick a track and a pace

Work down the track for your role in the roadmap. ★ shows how often a question comes up, from ★★★★★ (again and again) to ★☆☆☆☆ (rarely), and each track is ordered so that the top of the list pays off first.

Time you have	Research (RS / RE)	SWE	Applied AI / MLE	Student Researcher
About a week	stages 1–3	stages 1–2	stages 1–3	stage 1
Two to four weeks	stages 1–6	the whole track	the whole track	the whole track
More than that	add stage 7 and a design in your target team's area	add the ML quiz pages	add the research track's ML coding	add the research track's quiz pages

3. Practise one problem

Coding. Read only the Problem section. Before writing code, note what you would ask the interviewer (edge cases, tie-breaking, input sizes), then compare with the first lines of the reference solution, which list the points worth confirming. Give yourself a time limit and take the parts in order, treating each new part as a change of requirements: extend your code instead of starting over. Run it on the examples, then open the reference solution and compare approach, complexity and the # NOTE: comments. The checks call the functions named in the statement, so if you keep the same names and signatures you can usually run them against your own code.

Quiz. Set a timer of about three minutes per question and answer aloud, then write any derivation the question asks for on paper. Compare with the reference answer, which leads with what you would say and then shows the working; the checks recompute every number.

System design. Give yourself about 45 minutes and work through the prompt out loud or in a document: requirements, estimates, API and data model, architecture, then two or three deep dives. Then compare with the reference solution; the estimates are computed in the checks block, so you can change an input and rerun them.

Behavioral. Each page lists what the round asks, what each question probes and how a strong answer is structured, and ends with an outline to fill in with your own stories. Rehearse them aloud, and keep your filled-in version in my/, which git ignores.

4. Read a page

Every page starts with a table (type, priority, difficulty, roles, topics and, where known, format and round) and has exactly two sections:

Section	Coding	Quiz	System design	Behavioral
Problem	definitions, then one block per part: task, signature, example	numbered questions under topic headings	the system, its users, the numbers given, the scope	the round and its questions
Reference solution (collapsed)	per part: idea, derivation, code, with pitfalls as `# NOTE:` comments; then follow-ups and the checks	per question: the spoken answer, then the derivation; then the checks	requirements, data model and API, architecture, deep dives, follow-ups, estimate check	what is probed, how to structure the answer, an outline to fill in

Roles: RS research scientist · RE research engineer · SWE software engineer · MLE machine-learning engineer · Applied AI applied AI engineer · Intern internships and Student Researcher positions · PM product manager · TPM technical program manager.

5. Run the code

git clone https://github.com/Schuture/Google-DeepMind-Interview-Notes.git
cd Google-DeepMind-Interview-Notes
pip install -r requirements.txt                              # NumPy, SciPy, scikit-learn
python scripts/run_snippets.py coding/state-space-bfs        # one page
python scripts/run_snippets.py --all                         # every page

A page's ```python blocks run top to bottom as one script; ```py blocks are illustrative (bare signatures, code that contains planted bugs) and are not executed. CI uses Python 3.11.

Problems

Sorted by priority within each section. A dash under Difficulty means it has not been rated.

Coding (17)

#	Problem	Priority	Difficulty	Roles	Topics
1	Shortest Paths Through Locked Doors	★★★★★	Hard	SWE · RE · MLE · Intern	bfs, state-space-search, bitmask, path-counting, dijkstra, grid
2	Most Frequent Events in a Stream	★★★★☆	Medium	SWE · MLE · Applied AI · Intern	hash-map, sliding-window, frequency-counting, streaming, space-saving, heavy-hitters
3	Winning Positions in a Token-Moving Game	★★★★☆	Hard	SWE · RE · MLE · Intern	game-theory, dfs, retrograde-bfs, topological-order, sprague-grundy, graphs
4	Attention with a KV Cache and an Online Softmax	★★★★☆	Hard	RS · RE · MLE	attention, kv-cache, online-softmax, flash-attention, grouped-query-attention, numerical-stability
5	Debugging a Classifier That Does Not Learn	★★★★☆	Medium	RE · RS · MLE · Applied AI	debugging, softmax, data-shuffling, gradient-scaling, momentum, dropout, broadcasting, sanity-checks
6	Warm-Up: Scanning an Array and Traversing a Graph Depth-First	★★★★☆	Easy	SWE · MLE · Intern	arrays, binary-search, graph-construction, dfs, iterative-dfs, connected-components, cycle-detection
7	Python Generators, Unit Tests and Weighted Sampling	★★★☆☆	Medium	MLE · SWE · RE · Intern	generators, unit-testing, weighted-sampling, prefix-sums, binary-search, alias-method, chi-square-test
8	Code Review: Ranking the Defects in a Checkpointing Change	★★★☆☆	Medium	RE · SWE · MLE	code-review, checkpointing, atomic-writes, reproducibility, data-sharding, testing
9	Unit Conversions as a Weighted Graph	★★★☆☆	Medium	SWE · MLE · RE · Intern	graph, dfs, bfs, weighted-union-find, consistency-check, floating-point
10	Painting a Fence with the Fewest Strokes	★★★☆☆	Medium	SWE · MLE · Intern	greedy, divide-and-conquer, arrays, range-minimum, proof-of-optimality
11	Bit-Packing Encoders and Decoders	★★★☆☆	Medium	SWE · RE · Intern	bit-manipulation, varint, zigzag-encoding, frame-of-reference, serialisation
12	Snapshot Array: History, Compaction and Diffs	★★★☆☆	Medium	SWE · RE · MLE · Intern	hash-map, binary-search, versioning, memory-trade-offs, journaling
13	k-d Tree: Nearest Neighbours and Range Queries	★★★☆☆	Hard	RS · RE · SWE · MLE	kd-tree, nearest-neighbour, pruning, heap, range-search, curse-of-dimensionality
14	Backpropagation from Scratch	★★★☆☆	Medium	RS · RE · MLE · Intern	backpropagation, softmax-cross-entropy, gradient-check, initialisation, sgd-momentum
15	EM for a Gaussian Mixture: Derive and Implement	★★☆☆☆	Hard	RS · RE · MLE	expectation-maximisation, gaussian-mixture, log-sum-exp, jensen-inequality, k-means, bic
16	Focal Loss versus Cross-Entropy	★★☆☆☆	Medium	MLE · RS · RE · Applied AI	focal-loss, cross-entropy, class-imbalance, numerical-stability, initialisation, gradients
17	Binary Search on Bitonic Arrays and Monotone Functions	★★☆☆☆	Medium	RE · RS · SWE · MLE	binary-search, bitonic-array, exponential-search, monotone-functions, lower-bounds, bisection

Quiz (7)

#	Problem	Priority	Difficulty	Roles	Topics
1	ML Fundamentals: Metrics, Losses, Optimisers and Regularisation	★★★★★	Medium	RS · RE · MLE · Applied AI · Intern	precision-recall, roc-auc, cross-entropy, logistic-regression, adam, weight-decay, bias-variance, normalisation, huber-loss, gan, distribution-shift
2	LLM Fundamentals: Attention, Transformers and the Training Pipeline	★★★★★	Hard	RS · RE · MLE · Applied AI · Intern	attention, transformer, rope, kv-cache, scaling-laws, perplexity, tokenisation, post-training, sampling, mixture-of-experts
3	RL for Language Models: Policy Gradients, PPO, GRPO and DPO	★★★☆☆	Hard	RS · RE · MLE	policy-gradient, ppo, grpo, dpo, kl-regularisation, rlhf, reward-hacking, off-policy, importance-sampling
4	Maths Quiz: Probability, Statistics and Linear Algebra	★★★☆☆	Medium	RS · RE · MLE · Intern	bayes-theorem, expectation, markov-chains, maximum-likelihood, map-estimation, kl-divergence, svd, matrix-calculus, conditioning
5	Maths Quiz: Matrix Computation, Statistical Tests and Information Theory	★★★☆☆	Medium	RS · RE · MLE · Intern	matrix-multiplication, rank, matrix-inverse, pseudo-inverse, moments, central-limit-theorem, hypothesis-testing, chi-square-test, entropy, mutual-information, integration
6	CS Fundamentals Quiz: Memory, Concurrency and Floating Point	★★☆☆☆	Medium	SWE · Intern · RE	oop, memory-management, garbage-collection, race-conditions, deadlock, gil, cache-locality, floating-point, amortised-analysis
7	Code Comprehension: What Does This Code Compute?	★★☆☆☆	Medium	RS · RE · Intern · SWE	code-reading, convolution, padding, numerical-stability, streaming-statistics, attention-masks, reservoir-sampling

System design (8)

#	Problem	Priority	Difficulty	Roles	Topics
1	Design the Gemini App	★★★★☆	Hard	SWE · MLE · Applied AI	streaming, conversation-storage, context-management, model-routing, safety-filtering, multimodal-uploads, capacity-planning
2	Design an Evaluation System for an LLM Assistant	★★★★☆	Hard	RE · RS · MLE · Applied AI	evaluation, autoraters, side-by-side, bradley-terry, statistical-power, contamination, release-gating
3	ML Design: Predicting a Reaction Factor for Pairs of Molecules	★★★☆☆	Medium	MLE · RE · RS	eda, molecular-fingerprints, pairwise-models, symmetry, data-splitting, leakage, graph-neural-networks, active-learning
4	Design a Retrieval-Augmented Agent over Company Documents	★★★☆☆	Hard	Applied AI · MLE · SWE	rag, vector-index, hybrid-search, access-control, agents, prompt-injection, evaluation
5	Design a Recommendation System for a Content Feed	★★★☆☆	Medium	MLE · SWE · Applied AI	candidate-generation, two-tower, ranking, multi-task-learning, feedback-loops, cold-start, ndcg
6	Training a Model That Does Not Fit on One Accelerator	★★★☆☆	Hard	RE · MLE · RS	data-parallelism, tensor-parallelism, pipeline-parallelism, sharded-optimiser, checkpointing, fault-tolerance, loss-spikes
7	ML Design: A Data Flywheel for a Robot Manipulation Policy	★★☆☆☆	Hard	RE · RS · MLE	robot-learning, data-collection, dataset-curation, vision-language-action, evaluation-statistics, deployment-safety
8	ML Design: Predicting Which Data-Centre Machines Need Replacing	★★☆☆☆	Medium	MLE · RE · SWE	predictive-maintenance, label-construction, censoring, class-imbalance, categorical-embeddings, survival-analysis, precision-at-k, feedback-loops

Behavioral (5)

#	Problem	Priority	Difficulty	Roles	Topics
1	Recruiter Screen: Motivation, Research Interests and Logistics	★★★★★	—	All	why-gdm, motivation, research-interests, background, visa, logistics, compensation
2	Research Deep Dive: Paper Discussion and Research Talk	★★★★★	—	RS · RE · Intern	paper-deep-dive, research-talk, experimental-design, ablations, limitations, research-taste
3	Googleyness, Leadership and Culture: Competency Questions	★★★★★	—	All	star, why-gdm, collaboration, conflict, ownership, ambiguity, leadership, failure, responsibility
4	Team-Lead and Hiring-Manager Interviews: Experience, Research Taste and Fit	★★★★☆	—	RS · RE · SWE · MLE	research-taste, ml-experimentation, ramp-up, team-fit, open-ended-problems
5	Product Manager Loop: AI Product Sense and the AI Deep Dive	★★★☆☆	—	PM	product-sense, ai-product-strategy, agent-metrics, launch-risk, user-insights, ux-for-ai

Contributing

Corrections, new variants and translations are welcome, and so are accounts of recent interviews described in your own words. Found a wrong answer, a missing edge case, an outdated statement in the process guide or a sentence that does not make sense? Open an issue. For pull requests, see CONTRIBUTING.md.

License

The text (prose, tables and diagrams) is licensed under CC BY-NC 4.0; the code, both in the pages and under scripts/, under the MIT License.

Schuture/Google-DeepMind-Interview-Notes

Star growth

README