Coding questions by company

Cross-referenced from 1Point3Acres, Glassdoor, Blind, LeetCode discuss, candidate Medium write-ups, Hello Interview, interviewing.io, LinkJob, Tech Interview, Sundeep Teki guide. Updated May 2026.

Anthropic

Format: Take-home / async OA on CodeSignal (90 min, sometimes 60 live), 5-7 day window. Then onsite. 4-level progressive spec — must pass all tests at level N before moving to N+1. Very small known bank (~6 problems); no benefit from memorizing standard DSA. Onsite ML round is open-ended ("ML configuration system") plus a separate concurrency-flavored coding round. Reference checks during the loop. ~20-day timeline.

The "six known" Anthropic problems

1. Multithreaded web crawler — BFS from seed URL, same-domain filter, dedupe, #fragment handling. First sync, then ThreadPoolExecutor. Follow-ups: threads vs processes, politeness/robots.txt, distributed crawling.
2. In-memory key-value store / LRU cache — 4 levels: SET/GET/DELETE → filtered scans → TTL with timestamps → file compression + restart durability.
3. Stack-trace / sampling-profiler conversion — convert sampler stack snapshots to ordered start/end events. Handle recursion + identical consecutive stacks. Follow-ups: de-noising, N+ consecutive detection.
4. Distributed mode/median across N nodes — find statistical mode then median across 10 nodes with 10 B/s read, 1 B/s send/recv constraints.
5. Tokenize/detokenize round-trip — code review existing tokenizer, fix UNK handling, ensure invertibility.
6. Bank transaction system — multiple transaction types, progressive complexity.

ML round

• 55-min Google Meet in Colab on prompting + LLM engineering.
• ML configuration system design — open-ended.
• Open-ended ML design discussion.

Concurrency round

Always asked. See concurrency page. Common: thread-safe queue, fix race in shared cache, design rate limiter under contention.

Sources

• LinkJob Anthropic Q-Bank 2026
• Anqi Silvia — Anthropic concurrency Qs (2025)
• 1p3acres Anthropic
• Sundeep Teki AI Lab Guide

OpenAI

Format: CoderPad. 60–75 min phone screen. 4–6 hr final loop in 1–2 days. Coding bar is binary — they reject for 2/4 or low 3/4 even if ML/research/behavioral nail it. Each problem is multi-part; substantially more code than FAANG; less algorithmic, more "production." Decentralized hiring; behavioral round is non-standard.

Reported problems (2024–2026)

1. KV-store serialize/deserialize with arbitrary characters incl. delimiters (length-prefix encoding, Redis-style).
2. Time-based KV store (LC #981 flavor). Extensions: per-key locks vs global, disk persistence.
3. Spreadsheet API with getCell/setCell, dependency graph, cycle detection via DFS, optimize getCell to O(1).
4. GPU credit management — time-based credits with FIFO expiry, add/expire/consume.
5. Resumable iterator across nested structures with skip/reset.
6. In-memory DB with SQL-like ops (CREATE/INSERT/SELECT WHERE, joins).
7. Unix cd with symbolic-link resolution including cycle detection.
8. Multithreaded web crawler with rate limiting and dedup.
9. Toy interpreter (75-min RE round) — lex + parse + execute simple language with arithmetic, vars, control flow.

ML / research themes

Deep transformer architecture knowledge. KV-cache. Attention complexity. Batch norm. Regularization. Debugging transformer training (NaNs, masks, padding, optimizer state).

Sources

• Hello Interview — OpenAI 2025
• Anqi Silvia — 8 OpenAI questions (2025)
• interviewing.io OpenAI

Google DeepMind

Format: ~41-day average. PhD defense + rigorous engineering exam. AI tools prohibited in technical rounds (2026 policy). Undergraduate ML rapid-fire quiz that experienced PhDs frequently bomb (formal definitions of gradient flow, Bayesian inference, convergence proofs). Acceptance <1%.

Tracks

• Research Scientist: paper deep-dive (60 min) → research framing (60 min) → ML coding (60 min, unaided) implementing a piece of an ML pipeline (custom loss, attention, sampling routine, small training loop).
• Research Engineer: distributed training systems (Megatron/DeepSpeed/FSDP knowledge), eval harness design, general DSA.
• SWE: 2 medium-to-hard coding rounds (unaided), 1 system design, 1 domain depth.

Reported ML implementation tasks: custom loss from scratch, attention block, sampling/decoding code, small training loop in NumPy/PyTorch.

xAI (Grok)

Format: Engineer-led, very fast. Weight on written "exceptional work" statement. 20-min "hardest technical problem you've solved" presentation + Q&A defense. 45-60 min coding rounds, mix of algos and thread-safety/concurrency ("simple and correct beats clever and broken"). CodeSignal OA: 3 problems, 60 min.

Reported questions

• Implement a trie-based tokenizer
• Efficient attention for very long sequences (100K tokens) — design discussion
• Beam search with memory optimization
• CUDA kernel optimization for transformer inference
• Memory-efficient training algorithm for large models on limited GPU memory
• Distributed training system design for 100B+ param model
• Real-time inference system for Grok at 100K req/s
• Custom loss functions for LM pretraining
• Mixture-of-Experts architecture questions
• Concurrency-safe components (thread-safe queue/cache)

Cursor (Anysphere)

Format: TypeScript-first (Python OK for ML roles). AI tools permitted but interviewer judges your prompts. Sr/Staff: 4–8 hr take-home in their actual codebase. Heavy weight on craft round; fail mode is "passes coding but doesn't authentically use Cursor."

• Streaming Markdown parser in TypeScript (online test)
• Streaming edit application — apply incoming LLM tokens as real-time edits
• Multi-file diff/coordination tracking changes across many files
• Syntax-aware edit op
• File-tree diff modeling
• Build a hash tree in their codebase

Cohere

Format: Recruiter → HM (mostly projects) → virtual onsite: coding + ML concepts + system design + behavioral. Forward-deployed-platform-engineer flavor; expects you to debug a vector-DB upsert race condition rather than recite transformer math.

• Binary string reduction — count ops to reduce binary number to 0 (sub if odd / div by 2 if even)
• Streaming dedup — remove duplicate strings from stream without storing whole stream
• Longest unique substring (LC #3)
• System design: bit.ly URL shortener; real-time fraud detection with feature pipeline
• AI/ML: design knowledge-cutoff RAG mechanism; batch inference optimization (sub-batching with max-token + max-batch constraints, concurrent processing)

Mistral AI

Stages: LLM theory → coding → past project → tech manager → ML system design → take-home → values. Python coding round centers on refactoring rather than greenfield algorithms.

Perplexity AI

Format: Fast ~23-day loop, small bank, Python only.

1. Probability of each number appearing in a stream
2. Test if data stream is uniformly distributed (sample 3 random nums, verify uniform)
3. Substring before first stop word, then streaming version with memory constraint
4. Remove duplicate strings from stream → near-duplicates (case/punctuation/single-word diff)
5. LLM Provider Pool — Provider.query(prompt) + ProviderPool with fallback on failure across providers
6. CreditTracker class with add/subtract/check
7. Implement beam search given function signature + unit tests
8. Embedding model batch processing under max-batch + max-token constraints (concurrency)
9. System design: personal-finance multi-account sync
10. Kubernetes debugging — overloaded system, identify metrics

Cognition AI (Devin)

Heavy on customer-facing roleplay + pair programming + architecture deep-dive. Sample: "customer is angry that Devin deleted a critical file — walk through your response." "60-min onboarding workshop for senior engineers new to AI coding tools." Behavioral: "say no to a customer," "learn a new stack in days."

Sierra (Bret Taylor)

Replaced LeetCode with AI-native onsite: planning session (drive product ideation) + 2-hr building phase using AI tools/frameworks of choice. CoderPad multi-part, layered follow-ups; Python or TypeScript only. ~23-day loop.

Decagon

Recruiter → 60-min coding pair → 60-min system design → 60-min past project → behavioral.

• Implement n×n Tic-Tac-Toe game engine for two players
• Class that tracks conversation scores over a rolling time window
• LeetCode #84 (Largest Rectangle in Histogram)
• System design: AI agent system that resolves customer support tickets; knowledge ingestion + retrieval; eval framework for autonomous resolution quality

Apple (MLE / GenAI)

5–8 onsite rounds × 45–60 min. 1–2 coding rounds (LC medium-hard): rotate matrix 90°, merge K sorted lists, Mountain in array, LRU cache. Some rounds blend DSA + ML (e.g. implement cosine similarity, then DP extension). ML fundamentals: transformers, self-attention, RAG vs fine-tuning, embeddings, chunking.

Nvidia

ML system design + project depth. Coding includes stars pyramid, Maximum Product of Three Numbers (LC #628), producer-consumer ring buffer, parallel programming, matrix multiplication. ML: gradient descent, MLP/CNN/RNN/Transformer, Adam vs SGD, loss functions.

Waymo

CoderPad + Google Meet. Tech phone screen + ML design. Graph traversal, DP coding. ML: SGD variants, batch-size vs latency trade-offs, ML system design. 2 rounds in one day reported (1 coding + 1 ML design) Feb 2025.

Tesla Autopilot

HackerRank/Codility OA: medium algos in Python + simple ML tasks (feature engineering, eval). Onsite 3–5 rounds incl. ML system design for object detection pipeline (scalability, data pipeline, deployment).

Pinterest (MLE)

Tech phone screen: 3 ML fundamentals + 2 LC hard. Onsite: ML theory + LeetCode mediums. Topics: transformer arch, contrastive loss, learning-to-rank, vanishing gradients (which activations), recommender systems. Difficulty 3.4/5.

Snap (MLE)

1hr tech screen + 4hr onsite (2 LC rounds + ML system design + ML fundamentals). LC medium-hard. Sample: LC #4 Median of Two Sorted Arrays with O(log(m+n)) requirement. Task scheduling. Word dictionary. ML: explain attention, self-attention, cross-attention, multi-head attention; boosting; generative vs discriminative. ML infra design is differentiator.

Reddit (Senior MLE)

Recruiter → tech phone (build a model from provided data) → onsite (model design, feature engineering, applied ML in advertising). ML infra design as required round (feature stores, distributed training, online serving). Personalization/recsys focus.

Databricks (Staff MLE)

1hr CoderPad/Meet phone screen. LC medium-hard. ~20% pass rate. Onsite topics weighted toward graph algos, optimization, concurrency. LC tag composition: ~10 Easy / 19 Medium / 5 Hard; Array/HashTable heavy. Reference checks: 1 manager + 2 senior teammates, weighted heavily.

Snowflake

OA + 2hr tech phone screen (DSA + system design back-to-back). CoderPad. LC medium with hard follow-up. DP, BFS/DFS, binary search, linked lists; database internals twist. 30-min presentation interview on past project.

Notion (Backend SWE)

No LeetCode-style — practical problems only. Recently added AI-enabled interview requiring fluent use of Claude Code/Cursor.

Figma

3-5 rounds. LC medium DSA on coding round. System design × 2. Behavioral. Project. All coding questions "Figma-flavored."

Plaid

2 live coding questions on DS&A round. Backend/API/distributed-systems flavor. Real-world rather than pure LC.

Stripe

No LeetCode — practical, production-style problems on CoderPad. Sample: find pairs of transactions in 30s window summing to target; rate limiter for Stripe API; webhook signature validator with edge cases. ML coding round: dataset provided, build + evaluate model in 1 hour. System design: idempotency, exactly-once, distributed transactions, ledger flows, fraud, multi-currency. Integration round uses real Stripe API.

Airbnb (Staff MLE)

LC medium-hard, Airbnb-tagged. Classic DP problems (esp. on trees) reported. 4-5 stages. ML focus on dynamic pricing, trust signals, personalization.

Uber (Staff/MLE)

5-7 rounds over 4-8 weeks. 60/40 ML/coding split for MLE. Coding 1 (DSA) + Coding 2 (depth in specialization) for staff SWE. Recent 2025 SDE2 reports: DP on trees screening. ML system design: scalable pipelines, deployment, monitoring; offline vs online eval; A/B testing.

DoorDash (MLE)

60-min DSA round, 1 LC medium with 45 min for code+optimize. 22 reported problems, 6 Easy / 14 Medium / 2 Hard. Design HashMap, Longest Common Prefix, Jump Game. Graph traversal, scheduling, stateful services.

Anduril

4-6 onsite rounds. 2 LC easy + ML/CV concepts focus + ML design (harder). Heavy past-experience deep-dive. Perception roles need C++, PyTorch, TensorRT/ONNX deployment.

Cerebras

~26 day loop. 4 rounds. LC medium + parallel programming + matmul questions. Compiler engineer roles: MLIR, LLVM IR, parallelization/partitioning, novel program analysis.

Coinbase

OA: 2 math-heavy questions (DP + intervals/prefix sums). Onsite: 4 LC medium DSA in 90 min (target 3/4). Build something real (transaction management system, progressive requirements). System design: notifications with retries + queues.

Robinhood

OA: 4 questions (2 super easy + 1 hard + 1 medium). VO: 1 coding + 1 system design + 1 foundation.

AppLovin

5 rounds: 3 tech + 1 system design + 1 behavioral. LC medium, data-structure-design heavy. Classic: LRU cache O(1) then real-system extension (campaign config cache). Onsite: 2 questions starting easy with optimize follow-ups. ~3 weeks total.

Roblox

HackerRank OA + Roblox Assessment. 4 sections, ~2 hours. LC medium difficulty, often disguised in gaming/simulation context. 2 onsite tech rounds (Medium + Hard) + behavioral.

Crusoe / Lambda / Modal / Together / Baseten / MatX / Black Forest / World Labs / Physical Intelligence

Limited public data.

• Crusoe: 3 back-to-back (1 coding + 1 project review + 1 HM). Practical-oriented over LC.
• Lambda: SSH into EC2 + Linux/OOP probe + behavioral OOP (Python/Go). Take-homes for some roles.
• Modal: abstractions for distributed GPU workloads — system design heavy.
• World Labs / PI / BFL: expect VLA / world-model / 3D / SLAM / multimodal data-pipeline questions; PyTorch deployment; deep CV implementation.

Behavioral / culture themes per company

Key meta-takeaways

1. Frontier labs have moved away from pure LeetCode. 60–90 min progressive specs (Anthropic 4-level CodeSignal, OpenAI multi-part CoderPad) are the prototype. Drill these, not Hard sweeps.
2. "Implement multi-head attention from scratch in PyTorch" is the universal ML coding question (DeepMind, OpenAI, Anthropic, Mistral, Cohere, Perplexity, Snap, Apple). Drill until you can do it in 6 min cold.
3. Concurrency is the rising differentiator. Anthropic, xAI, OpenAI all probe explicitly. Don't skip.
4. AI tool policy varies wildly. DeepMind banned, Anthropic/OpenAI banned, Cursor/Sierra/Notion encouraged or required. Confirm per-company before the loop.
5. Take-homes are back at Sr/Staff: Anthropic OA, Cursor 4–8hr, Mistral, most agent startups.

Sources (by company)

• Anthropic: LinkJob, Anqi Silvia
• OpenAI: Hello Interview, Anqi Silvia
• DeepMind: Tech Interview 2026
• xAI: LinkJob, Exponent Dec 2025
• Cursor: Tech Interview
• Cohere: LinkJob
• Perplexity: LinkJob
• Decagon: Tech Interview
• Sierra: Sierra blog
• Cognition: Dataford
• Apple: Pranalibose 2025
• Stripe: interviewing.io
• Pinterest: Glassdoor
• Snap: Interview Query
• 1Point3Acres aggregator: 1p3acres
• Yuan Meng's MLE Interview 2.0: Yuan Meng
• Sundeep Teki frontier-lab guide: Sundeep Teki