Ex Google CEO: AI Can Create Deadly Viruses! If We See This, We Must Turn Off AI! - Eric Schmidt

KEYPOINTS

• 01

  If you are not using AI across every aspect of the business, Schmidt says you will not make it.

  Intro framing: he helped grow Google from roughly $100M to ~$180B scale. Adoption speed is strategy because AI will co-opt daily life through greater delight so gradually you may not notice control slipping. Productivity upside and autonomy risk arrive together—he insists leaders feel both. He pairs that urgency with the reminder that delight can hide how much of your world has already been co-opted—so dashboards should track dependency and control, not only efficiency gains.

• 02

  Biological misuse: viruses are relatively easy to make; truly bad designs are imaginable.

  He participates in a commission trying to stop that pathway. Non-proliferation sketch: a handful of labs in the US, China, and Britain might be manageable if tied together; if capability becomes trivial to copy, terrorists inherit a problem ‘not yet solved.’ China/Russia may lag one or two years on large models but will arrive. Title energy about ‘deadly viruses’ maps to this section: capability without custody is the nightmare, and he refuses to treat it as science fiction while still labeling parts as speculation.

• 03

  Sergey’s 70/20/10 rule—70% core, 20% adjacent, 10% other—generated tens of billions in extra profit over a decade.

  Core means search/ads; adjacent means bets like cloud; 10% funds new ideas institutionalized as Google X. An early X path produced Google Brain. The transferable operating principle: explicit exploration budget beats hoping spare evenings invent the next S-curve while AI compresses timelines. Anyone can ‘go do this,’ he says—meaning the portfolio math is portable to companies that are not Google, provided leadership protects the 10% from quarterly starvation.

• 04

  Risk-taking and fail-fast (Elon as exemplar) beat perfectionism when first mover captures disproportionate profit.

  Customers follow the right product, but the race is to arrive first. Iteration velocity and public willingness to be wrong sit beside capital as first principles for building something great.

• 05

  When asked his biggest AI fear clearly, he answers under-adoption—not adopting fast enough to fix shared problems.

  People want safety, healthcare, great schools; lagging tool use leaves those unsolved. An earlier ‘pull the plug’ fragment is cut by a mid-roll, so the coherent on-mic fear is slow adoption. Delight can still camouflage loss of control even while tools help—hold both sentences. Holding both fears—moving too slow on upsides and moving too loose on misuse—is the mature executive stance the episode keeps circling.

• 06

  AI enables ‘perfect’ misinformation; TikTok’s bandit algorithm is the addiction case study.

  Like a casino slot machine: keep serving what you engage, occasionally inject adjacent content—highly addictive. Synthetic media raises civic risk on a parallel track to biotech misuse.

• 07

  Open releases such as Meta’s Llama (~400B-class) spread capability to maximize core-business revenue.

  Open weights help builders and complicate non-proliferation metaphors. Zuckerberg’s founder-driven bet is hard for ordinary public companies to copy—and it changes who can fine-tune powerful systems.

• 08

  Teach analytical critical thinking; for an 18-year-old, add Python because models already write it.

  Path into judgment can be math, law, or entertainment—he cares that you think critically. Python is approachable and is the language AI systems use when coding for themselves; ‘make a game’ is the practical on-ramp. Tools churn; judgment compounds. Five or ten years ago blanket ‘learn to code’ advice felt different; he updates it: coding still helps, but critical thinking is the non-obsolete layer when models write Python for themselves.

• 09

  Automation should clear physically dangerous or soul-crushing repetitive work (security-guard example).

  Tired humans fall asleep on vigilance shifts; smarter systems can outperform them. Job disruption is real; his frame pushes workers and students toward AI fluency rather than nostalgia for dangerous drudgery.

• 10

  Culture and architecture outlive launches—Jobs closed vs Schmidt open battles included.

  He recounts philosophical fights with Steve Jobs on owning IP end-to-end versus more open approaches, plus Google’s user/safety focus in apps. Leadership lesson: culture compounds for decades after any single product moment.

• 11

  Five-year horizon: multiple more ‘turns of the crank’ on these systems—discontinuity comparable to nuclear in some minds.

  He invites listeners to imagine several more capability jumps and admits he has un-rosy moments about that future. Autonomy plus intelligence is what makes some observers rank AI risk beside or beyond nuclear metaphors—even while he simultaneously pushes faster beneficial adoption.