Skip to content
Phil's Library

Google DeepMind interview with Rupert Kemp

Role: act as “CEO proxy for complex, ambiguous, high-stakes AI programs”

  • OpenAI / Microsoft partnership

  • AI Supercomputer program

  • Cross-company AI forums (when reasoning models first emerged) → drive ambition (educate & showcase greatness), build the platform, differentiate

  • Where am I successful?

    • As an execution lever for CEO, operating at the interface of research and product
      • Product: I get products done (GitHub Copilot, XNA) and design teams that get work done (Frontier Impact)
      • Executives: I speak SLT and researcher (OpenAI → MSFT, MSR → MSFT, Supercomputer program, AI for Science)
      • Builder: I build my own, get hands-on

  • Why DeepMind and this team?

    • Want to work in an end-to-end organization that sees AGI as a systems problem, integrated with society
    • Work in a small team with high leverage and influence.

  • Have a Theory of Success

    • leaf nodes / foundation for reality
    • Stop thinking you can predict the future → you can predict general trends, but have to jump in
    • Hazards of prophecy: failure of nerve, failure of imagination

  • Innovation is all about Ambidextrous Leadership:

    • The best way to have good ideas is to have lots of ideas - speak to people!
    • “No prize for pessimism”
    • Be clear about whether you’re doing something for prestige or profit - helps you be honest about whether you’re looking for knowledge/capability or commercial evidence
    • Drive the cost of experimentation to zero
    • Make sure there’s high psychological safety - it’s about good experiments, not hitting home runs every time (if you are, you’re not being ambitious enough!)
    • Remember it’s not a single-player, static game
    • Don’t get hung up on small costs, and move quickly when you have a ‘holy shit’ moment

  • Great partnerships…

    1. Build trust – navigate future disagreements or issues
    2. Maintain alignment – defining and revisiting shared goals and desired outcomes often
    3. Show persistence – tackle complex problems together
    4. Have fearlessness – don’t be afraid to do hard things or take risks

Productizing AI

Section titled “Productizing AI”
  • GITHUB COPILOT, AI for Science and Frontier Impact, AI Early Adopter program (enterprise knowledge graph)
  • research → product
  • capability → user value (replace workflows!, drive use, drive data)
    • identify what’s possible
    • user problem first, not model
    • latency constraints?
    • safety & hallucination risks
    • UX
    • iteration loops (evals → deploy → feedback)
  • Example questions:
    • “Take a recent AI research breakthrough — how would you turn it into a product?” → Mariner agentic OS? Use Chromebooks?
    • “How do you decide if a model is ready to ship?”
    • “What makes an AI product actually useful vs just impressive?”

CEO-level Prioritization

Section titled “CEO-level Prioritization”
  • Core tenets ahead of time. What has the best leverage?
  • Judgment under uncertainty → identify core tenets ahead of time
  • Ability to think at portfolio level
    • AGI trajectory
    • strategic leverage
    • time-to-impact vs long-term value
  • Example questions:
    • “You have 3 major AI initiatives—how do you prioritize?”
    • “What should DeepMind focus on next beyond Gemini?”
    • “If you were advising the CEO, what would you deprioritize?”

Cross-functional Conflict

Section titled “Cross-functional Conflict”
  • AI Supercomputer program → work on alignment, work outside the boundaries. DISAGREEMENT IS OK !
  • Executive maturity
  • Conflict navigation
    • don’t “choose a side” immediately
    • creates: shared metrics, decision frameworks
    • escalate only when necessary
  • Example questions:
    • “Describe a difficult stakeholder situation” (Glassdoor)
    • “Research wants to delay release; product wants to ship—what do you do?”
    • “Two senior leaders disagree—how do you resolve it?”
    • “How do you influence without authority?”

Ambiguous Problem Solving

Section titled “Ambiguous Problem Solving”
  • Structure under ambiguity
    • Breaks problem into: goals, constraints, unknowns
    • Explicitly states assumptions
  • Example questions:
    • “How do you approach a problem with no clear solution?”
    • “How do you navigate new information?” Identify sources of uncertainty
    • “What would you do in this situation?”

Applied AI Depth

Section titled “Applied AI Depth”

  • “What are the biggest bottlenecks in deploying AI products today?”

  • “What’s hard about evaluation in LLM systems?”

  • “Where do current models fail?”

  • hallucinations

  • eval difficulty

  • distribution shift

  • infra + latency

CEO Communication & Influence

Section titled “CEO Communication & Influence”
  • “How would you present a controversial recommendation to the CEO?”
  • “How do you get alignment across senior stakeholders?”
  • “Tell me about a time you influenced executives”

Behavioral

Section titled “Behavioral”
  • “Why this role?”
  • “Tell me about a failure”
  • “Tell me about a time something went wrong”

CASE 1: “Research → Product”

Section titled “CASE 1: “Research → Product””
  • “DeepMind has a new model capability. How do you turn it into a product in 6 months?”
    • roadmap thinking & tradeoffs
      1. Define user value
      2. Select use case (narrow first) and stakeholders / delivery vehicle
      3. Identify constraints (latency, safety)
      4. MVP scope
      5. Iterate

Case 2: “Ship Vs Safety”

Section titled “Case 2: “Ship Vs Safety””
  • “Model is powerful but risky. Do you launch?” - staged rollout - eval thresholds - guardrails - NOT binary yes/no

Case 3: “CEO prioritization”

Section titled “Case 3: “CEO prioritization””
  • “Where should DeepMind invest next?”
    • on the AGI path
    • competitive landscape
    • product leverage

Questions YOU Should Ask Them

Section titled “Questions YOU Should Ask Them”
  • “How does the CEO prioritize across research vs product vs safety?”
  • “What does success look like in this role after 12 months?”
  • “Where are the biggest execution bottlenecks today?”
  • “How do you decide what not to work on?”

AREAS OF AI I CARE ABOUT

Section titled “AREAS OF AI I CARE ABOUT”
  • Network effects of AI systems (memory, gets better as others use it)
  • Role of open source models (Gemma 4)
  • Role of fine-tuning (specialisation vs. generalisation)
  • AI operating system (Project Mariner)
  • Embodied systems and AI in the physical world (Gemini Robotics, Age of Experience / David Silver)
  • AI for Science (AlphaFold|Genome|Missense|Earth / Weather / Math:Evolve/Proof/Geometry, Fusion)
  • ALL WITH SAFETY !

—

How to Run an Innovation Pipeline

Section titled “How to Run an Innovation Pipeline”

An innovation pipeline requires a disciplined, evidence-based, data-driven process for connecting innovation activities into an accountable system that rapidly delivers solutions to hard problems.

—

Building Great Teams

Section titled “Building Great Teams”
  • Pioneers: explore new concepts, the uncharted land. Show wonder but fail.
  • Settlers: turn the half baked thing into something useful for a larger audience. They build trust. They build understanding. They make the possible future actually happen.
  • Town Planners: take something and industrialise it taking advantage of economies of scale.

What you want is brilliant people in each of these roles.