OpenAI MLE Interview Questions (2026)

Questions reported from OpenAI machine learning engineer interviews. This bank tracks 35 questions asked at OpenAI (16 with free model answers). Practice them live with DeepOffer's AI voice interviewer.

← All 465 questions

Practice this question with an AI interviewer

Get asked follow-ups live, then receive a scored report — like a real MLE interview loop.

Start AI mock interview

#27 Derive the Main Steps and Complexity of Self-Attention
#40 Architect a Production Service for Autoregressive LLM Generation ★
#41 How Would You Train Logistic Regression Using Gradient Descent from Scratch? ★
#43 Describe the KV cache, derive its memory formula, and discuss what it costs at production scale. ★
#59 Explain self-attention, its computational complexity, and your options as contexts grow long.
#60 Clarify how cross-entropy, KL divergence, and perplexity relate, and why cross-entropy is the standard LM training loss.
#77 Given D=4096, F=4D, V=32000, L=64, compute a transformer's parameter count and the fraction that are attention parameters.
#90 Give a decision framework for choosing between prompting, RAG, and fine-tuning, with cost and latency attached. ★
#122 Compute a KL divergence given different random variables.
#123 If a classifier's accuracy is 1, what are the lower and upper bounds on the loss for a single training example?
#124 Given two models at 85% and 82% accuracy, explain which you would pick and why.
#125 Describe how you handle missing data in Pandas.
#177 Your p99 latency doubled after a deploy with no model change — walk through the diagnosis. ★
#188 Design the serving stack for a ChatGPT-scale consumer assistant: hundreds of millions of weekly users, streaming chat, multiple model tiers. ★
#221 Design an enterprise RAG assistant over 10M documents with per-user permissions. ★
#238 Describe how to make a production tool-calling agent loop reliable enough to ship.
#240 Design human-in-the-loop approval for an agent that takes consequential actions. ★
#243 Explain how to detect and measure hallucinations in a production RAG system. ★
#248 A customer says "the model got worse" after a model-version upgrade — how do you verify and respond?
#294 Implement a BPE tokenizer from scratch: train merge rules, then encode and decode without external libraries. ★
#295 Implement scaled dot-product self-attention in NumPy or PyTorch with an optional causal mask. ★
#304 Debug a Transformer: find the classic bugs. ★
#309 Design and implement an in-memory key-value store supporting set plus transactional begin, commit, and abort. ★
#312 Implement an LRU cache with O(1) get/put, then add TTL. ★
#314 Implement a token-bucket rate limiter, then make it distributed. ★
#338 Simulate plant infection / cellular automaton on a grid: compute days to stabilize, an immune-cell variant, and an immunity countdown (5 progressive parts).
#339 Implement GPU credit operations (create_grant / subtract / get_balance) with expiry priority and out-of-order operations, then optimize for production.
#340 Build a text editor in 4 parts: buffer ops, dual-stack undo/redo, AI autocomplete with Trie + word frequencies, and real-time collaboration/OT.
#341 Schedule data-labeling tasks across (task, model) and (task, human) pairs.
#353 Refactor ~120 lines of working but messy code with passing tests, improving the architecture without breaking them.
#363 Describe a time you made a mistake. ★
#368 Explain your views on AGI and AI safety.
#369 Explain why you want to work at OpenAI specifically.
#417 Design an image-generation service (Midjourney/DALL-E style): queue management, GPU scheduling, and content-safety filtering at scale
#418 Design an autonomous travel-booking agent: orchestrator, tool gateway, credential management, rollback, and risk-classified actions