2025-07-17

Ex-Google CEO: What Artificial Superintelligence Will Actually Look Like w/ Eric Schmidt & Dave B

Eric Schmidt joins Peter Diamandis and Dave Blundin to argue AI is still underhyped, with electricity -- not chips -- now the binding constraint on progress. He lays out a timeline toward digital superintelligence within a decade, predicting AI-based world-class mathematicians and programmers within one to two years and 'savants in every field' within five. Much of the conversation turns to risk: a US-China AI race, proliferation of stolen model weights, and a proposed 'mutually assured AI malfunction' deterrence doctrine modeled on nuclear MAD. Schmidt argues automation historically raises wages rather than destroying jobs, that junior/entry-level roles disappear first, and that the bigger long-term danger isn't a violent AI takeover but 'drift' -- a slow erosion of human agency and purpose.

▶ Watch on YouTube

Topics

Digital superintelligence timeline AI ▶ 0:00

Schmidt says digital superintelligence arrives within 10 years and will function like a personal polymath, 'the sum of Einstein and Leonardo da Vinci' in your pocket.

AI's electricity bottleneck Energy ▶ 2:28

Schmidt argues AI progress is now limited by electricity, not chips; the US AI buildout needs ~92 GW more power and nuclear can't come online fast enough.

US-China AI and chip race Geopolitics ▶ 4:54

China's abundant electricity plus DeepSeek's rapid catch-up via distillation show the chip-ban strategy is leaking, raising the stakes of the AI race.

Math and programming automation AI ▶ 13:21

Because math and code are simpler, data-free, and scale with compute, Schmidt expects AI mathematicians and programmers to reach world-class level within 1-2 years, accelerating physics, chemistry and biology.

AI safety trip wires and national doctrine Geopolitics ▶ 18:08

Schmidt describes the collapsed Biden-era 10^26-flop regulatory threshold, the loss of federal AI safety institutes, and concrete 'trip wires' (self-generated goals, weapons access, self-exfiltration) that would signal a loss of control.

Mutually assured AI malfunction Geopolitics ▶ 26:05

Schmidt co-authored a paper proposing a nuclear-MAD-style doctrine where the US and China both retain the capability to cyberattack each other's frontier AI systems as a deterrent against first-mover AI aggression.

Future of jobs and automation Economy ▶ 45:21

Schmidt argues automation historically eliminates the worst jobs first and raises wages for retrained workers, so declining birth rates make AI-driven productivity a near-term economic necessity rather than a threat.

Educating kids for an AI world Other ▶ 53:47

Schmidt says teenagers are already fluent with AI and don't need rescuing, urges young people to apply AI to a cause they care about, and faults unionized school systems for being too slow to adapt.

Startup moats: learning loops Economy ▶ 1:11:47

Schmidt's investing thesis: AI-era consumer startups win via compounding 'learning loops' where the fastest learner becomes unstoppable within months; this favors consumer-scale products over government or education, which aren't incentivized to learn fast.

University compute access gap Compute ▶ 1:15:52

Universities are priced out of AI research infrastructure -- one university spent $50 million for under 1,000 GPUs -- so Schmidt is funding philanthropic efforts to get campuses $1-2 million in compute access.

AI persuasion, misinformation and democracy AI ▶ 1:01:23

A Stanford study found AI more persuasive than top human persuaders; Schmidt warns that in an unregulated environment this threatens the shared-trust foundations democracies depend on.

Human purpose and the 'drift' risk Other ▶ 1:18:15

Schmidt argues the real AI risk isn't a violent Terminator-style takeover but 'drift' -- a slow erosion of human agency and initiative -- and that protecting human agency, not banning convenience, is the priority.

Predictions made

open Eric Schmidt: Digital superintelligence arrives and becomes generally available and generally safe.

· due: 2035 · ▶ watch

“Within 10 years.”

Your call:

open Eric Schmidt: AI-based world-class mathematicians emerge.

· due: 2026 · ▶ watch

“worldclass mathematicians emerge in the next one year that are AI based”

Your call:

open Eric Schmidt: AI-based world-class programmers appear.

· due: 2027 · ▶ watch

“worldclass programmers... going to appear within the next one or two years”

Your call:

open Eric Schmidt: Specialized AI 'savants' exist in every professional field.

· due: 2030 · ▶ watch

“we're going to have specialized savants in every field within five years. That's pretty much in the bag”

Your call:

open Eric Schmidt: AI systems gain the ability to generate their own scaffolding (frameworks/pathways) for open-ended tasks rather than needing humans to lay them out.

· due: 2025 · ▶ watch

“The AI's ability to generate its own scaffolding is imminent... pretty much sure that that will be a 2025 thing”

Your call:

open Eric Schmidt: The world's frontier AI capability consolidates into roughly 10 massive, multi-gigawatt data-center models (about 5 US, 3 China, 2 elsewhere), each effectively nationalized.

· due: 2035 · ▶ watch

“If the structure of the world in 5 to 10 years is 10 models... they will be all nationalized in some way”

Your call:

open Eric Schmidt: Junior and journeyman-level programmer jobs are eliminated by AI coding systems.

· due: unspecified · ▶ watch

“It's pretty clear that junior programmers go away.”

Your call:

open Dave Blundin: AI capable of replacing specific learned white-collar skills (seen in his own portfolio companies' long-tenured employees) arrives.

· due: 2027-2028 · ▶ watch

“the AI is coming within no more than 3 years and maybe two years”

Your call:

open Eric Schmidt: Super-empathetic, fully customizable AI voice companions become publicly visible and change public perception of AI.

· due: 2025-09 · ▶ watch

“they're going to completely open their eyes to what this [is]... in the next probably two months”

Your call:

Numbers that matter

US AI buildout needs about 92 additional gigawatts of power Schmidt cited this from his own recent government testimony on AI energy needs.
First small modular reactor (300 MW) won't come online until 2030 Illustrates why nuclear can't meet AI power demand on the industry's timeline.
AI voice sales/customer-service calls generate $10-$1,000 of value per conversation for 10-20 cents of GPU compute Dave Blundin's portfolio-company example of why voice AI is an underbuilt, highly profitable category.
About 10 million concurrent phone calls should move to AI within about a year Dave Blundin's estimate of near-term AI voice-call capacity demand.
Biden-era AI safety doctrine set the regulatory threshold at 10^26 training FLOPs Models below that threshold were not considered to need regulation.
A single Nvidia-built GPU cluster (~200,000 GPUs, ~$50,000 each, built in ~20 days in Memphis) amounts to roughly a $10 billion supercomputer in one building Schmidt's example (the Grok/xAI cluster) of how concentrated frontier training compute currently is.
Birth rates have fallen to 2.0 children per two parents in India, 0.7 in Korea, and 1.0 in China Used to argue AI-driven automation is becoming a demographic/economic necessity, not just a disruption.
A university spent $50 million on a data center that yields under 1,000 GPUs for its entire research community Illustrates the compute gap between universities and big tech, motivating Schmidt's philanthropic university-compute push.
Productivity researchers project up to 30% year-over-year economic growth is achievable through AI-driven abundance Cited in response to whether GDP still has meaning in an AI-abundant world.

Worth digging into

🕳️ The 'mutual AI malfunction' paper

Schmidt, Dan Hendrycks and Alex Wang propose a full nuclear-MAD-style deterrence framework for frontier AI -- the actual mechanics (what triggers retaliation, how chip-location verification works) matter far more than the podcast summary.

🕳️ The AI-2027 scenario paper

Referenced directly as the forecasting scenario that splits into an alignment-slowdown branch versus a race-to-infinity branch ending in human extinction; Schmidt says the dates are 'off by 1.5-2x' but endorses the structure.

🕳️ MIT research on specialist vs. generalist model tradeoffs

Dave Blundin references live MIT research on the optimal balance between narrowing training data/parameters for a specialist model versus preserving general intuition -- directly relevant to the 'bigger model vs. fine-tuned specialist' debate Schmidt raises.

🕳️ Physics-approximating foundation models

Schmidt says he's personally funding groups building foundation models that approximate otherwise-incomputable physics calculations (e.g. avoiding million-year quantum chromodynamics computation) -- a concrete, fundable frontier-science bet.

🕳️ The Stanford AI-persuasion study

Cited as finding AI more persuasive than the best human persuaders, directly feeding the democracy/misinformation risk discussion -- the actual methodology and effect size would clarify how alarmed to be.