Google and Google DeepMind

English · Русский · AI Interview Atlas › Companies

Big Tech · International · Loop reviewed 2026-09-26

Guides describe engineering-management and Gemini product interviews. One L6 management candidate reports code review and debugging. These sources do not establish a general software- or research-engineer loop.

On this page: Interview loop · Questions (18)

Roles covered

• AI Leadership: Engineering manager, AI product manager (senior to group)

Interview loop

Stage	What happens	Basis
Engineering manager path	Recruiter screen, a technical phone screen with coding, then an onsite with coding, system design and leadership interviews; a hiring committee reviews the feedback and team matching follows.	† prep guide or compilation Google Engineering Manager (EM) Interview Guide, Aced (formerly Exponent), retrieved 2026-09-26
Engineering manager report (L6)	Googleyness and behavioral questions (a team disagreement, attrition during a reorganisation), two system design rounds pushed down to implementation detail, an EM round on low performers, and a must-pass code review and debugging round on 200+ lines in a shared document without an IDE.	🗣 candidate report Google Engineering Manager (L6) Interview Experience, Aced (formerly Exponent), candidate report, retrieved 2026-10-05
Gemini product manager path	Recruiter and hiring manager screens on AI-first thinking, a product design round with about 15 minutes of vibe-coding a prototype, an execution and strategy round on debugging LLM systems, a generative-AI system design round and a leadership round.	† prep guide or compilation Gemini Product Manager Interview Guide, Aced (formerly Exponent), retrieved 2026-09-26

Coding: varies by role. One EM candidate reports code review/debugging; AI prototyping for PMs is a secondary-guide claim. Confirm your role's format. (Google Engineering Manager (EM) Interview Guide, Aced (formerly Exponent), retrieved 2026-09-26, Google Engineering Manager (L6) Interview Experience, Aced (formerly Exponent), candidate report, retrieved 2026-10-05, Gemini Product Manager Interview Guide, Aced (formerly Exponent), retrieved 2026-09-26)

Questions reported for Google and Google DeepMind

LLM fundamentals

• Compare greedy decoding, beam search and temperature, top-k and nucleus sampling; give a failure mode for each.
  • Knowledge · Asked at: Google and Google DeepMind †, Perplexity †
  • Tests: Whether probability manipulation is connected to output diversity and quality.
  • A strong answer covers:
    • Describe each as a transform: argmax, approximate sequence-likelihood search, logit division by T, top-k truncation, cumulative-mass truncation.
    • Name failures: greedy and beam repeat or go bland, high temperature derails, top-k ignores distribution shape, top-p admits flat tails.
    • Choose by task, deterministic for extraction and sampled for generation, and measure accuracy together with diversity across seeds.
  • Read: The Curious Case of Neural Text Degeneration (Holtzman et al., arXiv)

Inference, serving and cost

• Compare tensor, pipeline, data, sequence and expert parallelism for a large model deployment.
  • Knowledge · Asked at: Amazon †, Google and Google DeepMind †, Meta †
  • Tests: Whether communication patterns and memory limits determine the choice.
  • A strong answer covers:
    • Name each split and its traffic: TP per-layer all-reduce, PP stage activations, SP key-value exchange, EP token all-to-all.
    • Fit weights and KV first: TP inside the NVLink domain, PP across nodes, then traffic-free data-parallel replicas for throughput.
    • Measure per-GPU memory headroom, communication share of step time, pipeline bubbles and expert load imbalance at target latency.
  • Read: vLLM documentation (vLLM project) · Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity (Fedus et al., arXiv) · Efficiently Scaling Transformer Inference (Pope et al., arXiv) · Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism (Shoeybi et al., arXiv)

Fine-tuning and post-training

• A long pretraining run suddenly diverges. How do you investigate data, numerics and optimiser state before resuming from a checkpoint?
  • Applied scenario · Asked at: Google and Google DeepMind †
  • Tests: Whether recovery is based on a reproducible diagnosis.
  • A strong answer covers:
    • Locate the first anomalous step in gradient norm, attention-logit and activation magnitudes, and update-to-weight ratios.
    • Replay from the prior checkpoint with identical data order; a reproduced spike implicates data or numerics, otherwise suspect hardware.
    • Apply one targeted fix, such as skipping batches, a lower learning rate, tighter clipping or z-loss, and compare the replayed trajectory.
  • Read: PaLM: Scaling Language Modeling with Pathways (Chowdhery et al., arXiv) · Small-scale proxies for large-scale Transformer training instabilities (Wortsman et al., arXiv)
• Design training for a model that cannot fit on one accelerator; explain how you partition state and keep communication affordable.
  • System design · Asked at: Google and Google DeepMind † · ✍ Answer
  • Tests: Whether model state, activations and network topology determine the parallel plan.
  • A strong answer covers:
    • Separate parameters, gradients, optimiser state and activations before choosing a partitioning plan.
    • Explain ZeRO state sharding and its communication trade-offs; distinguish it from splitting layer computation.
    • Measure peak memory, communication and throughput on the target topology before scaling the run.
  • Read: ZeRO: Memory Optimizations Toward Training Trillion Parameter Models (Rajbhandari et al., arXiv)

Evaluation and observability

• Build the evaluation harness for a new model release, covering reproducibility, execution isolation and comparable results.
  • System design · Asked at: Google and Google DeepMind †
  • Tests: Whether evaluation infrastructure can support trustworthy comparisons.
  • A strong answer covers:
    • Pin everything that shapes a score: dataset version, prompt template, decoding settings, grader and seeds.
    • Run model-generated code in isolated, network-restricted sandboxes with resource limits and clean state per task.
    • Report means over repeated samples with confidence intervals, and rerun a baseline model to detect harness drift.
  • Read: Demystifying evals for AI agents (Anthropic) · Evaluation best practices (OpenAI API documentation)

AI system design

• Design a streaming chat service for hundreds of millions of users, including capacity, conversation storage and graceful overload behaviour.
  • System design · Asked at: Anthropic †, Google and Google DeepMind †, Meta †, OpenAI † · ✍ Answer
  • Tests: Whether scale assumptions lead to explicit capacity and reliability decisions.
  • A strong answer covers:
    • Estimate concurrent streams, token demand and storage from declared traffic assumptions.
    • Separate admission control, inference scheduling and durable conversation state.
    • Plan capacity headroom, cancellation, overload rejection and recovery from partial failure.
  • Read: Site Reliability Engineering (Google) · Efficient Memory Management for Large Language Model Serving with PagedAttention (Kwon et al., arXiv)

Practical coding

• Implement scaled dot-product attention with a causal mask and tests that would expose attention to future tokens.
  • Coding · Asked at: Amazon †, Anthropic †, Google and Google DeepMind † · ✍ Answer
  • Tests: Whether shapes, masking and numerical stability are correct in executable code.
  • A strong answer covers:
    • Check batch, sequence and head dimensions before implementing score multiplication.
    • Use stable softmax after masking disallowed positions.
    • Test small hand-computable cases and ensure changing future inputs cannot change earlier outputs.
  • Read: Attention Is All You Need (Vaswani et al., arXiv)
• Implement multi-head attention, then modify the implementation so several query heads share each pair of key and value heads.
  • Coding · Asked at: Google and Google DeepMind †
  • Tests: Whether tensor layout and head grouping are implemented consistently.
  • A strong answer covers:
    • Fix shapes first: batch, sequence, heads and head dimension; require query heads divisible by KV heads.
    • Map query head h to KV head h // group_size; when expanding, use repeat_interleave, not tile.
    • Test that equal head counts reproduce multi-head attention and a single KV head gives multi-query attention.
  • Read: GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints (Ainslie et al., arXiv) · Attention Is All You Need (Vaswani et al., arXiv)
• Implement temperature scaling, top-k filtering and nucleus sampling for a logits vector, including boundary cases.
  • Coding · Asked at: Google and Google DeepMind †
  • Tests: Whether probability mass, ties and invalid parameters are handled correctly.
  • A strong answer covers:
    • Define the parameter contract: temperature zero as argmax, valid ranges for k and p, non-finite logits.
    • Use a stable softmax; nucleus keeps the smallest sorted prefix reaching mass p, an O(V log V) sort.
    • Test ties at the cutoff, p near one, k above vocabulary size, and frequencies under a fixed seed.
  • Read: The Curious Case of Neural Text Degeneration (Holtzman et al., arXiv)

AI product strategy and metrics

• How do you decide whether an LLM is useful for a proposed product problem?
  • Product case · Senior · Asked at: Google and Google DeepMind † · ✍ Answer
  • Tests: Whether you start with the task and evidence rather than the technology.
  • A strong answer covers:
    • Describe the job, baseline solution and cost of mistakes.
    • Check data access, evaluability, latency and the need for predictable behaviour.
    • Run a bounded comparison and name a reason to reject the LLM approach.
  • Read: Rules of Machine Learning (Google for Developers) · Building Effective AI Agents (Anthropic)
• Design a Gemini-based assistant for university students.
  • Product case · Senior · Asked at: Google and Google DeepMind † · ✍ Answer
  • Tests: Whether a clear user need drives scope, learning outcomes and safeguards.
  • A strong answer covers:
    • Pick a student segment and one recurring unmet need.
    • Design a usable flow with uncertainty, verification and misuse in mind.
    • Measure task success and learning value alongside retention.
  • Read: Measuring the User Experience on a Large Scale: User-Centered Metrics for Web Applications (Rodden, Hutchinson and Fu, Google Research) · AI Risk Management Framework (NIST)
• Users say Gemini gives incorrect answers with excessive confidence. What would you change?
  • Product case · Senior · Asked at: Google and Google DeepMind † · ✍ Answer
  • Tests: Whether you diagnose failure types and improve calibrated user trust.
  • A strong answer covers:
    • Collect representative failures and separate knowledge, retrieval and presentation issues.
    • Compare grounding, abstention and interaction changes on the same tasks.
    • Track unsupported claims, useful task completion and segment-level regressions.
  • Read: AI Risk Management Framework (NIST) · Demystifying evals for AI agents (Anthropic)

Leading engineering teams

• How would you handle an engineer who is not meeting performance expectations?
  • Applied scenario · Senior · Asked at: Google and Google DeepMind 🗣 · ✍ Answer
  • Tests: Whether you diagnose a performance gap, provide fair support and make expectations and follow-up explicit.
  • A strong answer covers:
    • Establish observable expectations and examples of the gap; check role clarity and working conditions.
    • Hear the engineer’s account, agree support and define measurable improvement with review dates.
    • Follow the organisation’s people process, distinguish improvement from reassignment or exit, and protect confidentiality.
  • Read: [re:Work] Google's New Manager Training Slides (Google re:Work)
• How would you handle a high-performing engineer whose behaviour repeatedly creates conflict with teammates?
  • Applied scenario · Senior · Asked at: Google and Google DeepMind 🗣 · ✍ Answer
  • Tests: Whether you address observable collaboration problems without exempting high performers or deciding the case before hearing those involved.
  • A strong answer covers:
    • Establish specific behaviour and team impact; hear different accounts without public blame.
    • Explain that collaboration is part of performance and agree observable changes and support.
    • Review both delivery and collaboration; use the established people process if harmful behaviour persists.
  • Read: [re:Work] Google's New Manager Training Slides (Google re:Work)
• Tell me about a difficult organisational change you managed that involved high-performing engineers leaving. How did you respond?
  • Behavioral · Senior · Asked at: Google and Google DeepMind 🗣 · ✍ Answer
  • Tests: Whether you can explain actual attrition, your responsibility and the effect of your response on the remaining team.
  • A strong answer covers:
    • Use a real change and actual departures; state your authority and distinguish known reasons from inference.
    • Explain how remaining engineers reacted and what you changed in workload, communication or management practice.
    • Give follow-up evidence, limits of attribution and what you would handle differently.
  • Read: Organizational Change Management (GitLab Handbook)
• How would you stabilise an engineering team after a change in management?
  • Applied scenario · Senior · Asked at: Google and Google DeepMind 🗣 · ✍ Answer
  • Tests: Whether you establish trust and operating clarity before imposing broad changes.
  • A strong answer covers:
    • Listen to team members and partners and identify urgent delivery or wellbeing risks.
    • Clarify priorities, decision ownership and communication routines.
    • Make a small number of justified changes and check their effect.
  • Read: Organizational Change Management (GitLab Handbook)
• Describe a team disagreement that you resolved without discarding the valid concerns on either side.
  • Behavioral · Senior · Asked at: Google and Google DeepMind 🗣 · ✍ Answer
  • Tests: Whether you uncover underlying needs instead of forcing a superficial compromise.
  • A strong answer covers:
    • Pick a disagreement where both positions protected something real, and state each in its strongest form.
    • Show how you moved from positions to the needs behind them and found an option meeting both.
    • Close with what was decided, whether both concerns stayed addressed, and how you checked.
  • Read: [re:Work] Google's New Manager Training Slides (Google re:Work)

Behavioral and values

• Describe a disagreement with a researcher or technical lead about priorities and what followed.
  • Behavioral · Senior · Asked at: Google and Google DeepMind †
  • Tests: Whether you can reason across research and engineering incentives.
  • A strong answer covers:
    • Pick a dispute over priorities where research and engineering goals genuinely pulled in different directions.
    • Explain what each side was rewarded for — new findings or reliable delivery — and what each choice cost.
    • Show how it was settled, for example by a time-boxed experiment, and what followed for both sides.
  • Read: Design Docs at Google (Industrial Empathy) · Communication (GitLab Handbook)

← Canva · Meta →