2025-11-07

Part 1: Eric Schmidt and Fei-Fei Li: Human Life After Artificial Superintelligence | EP #206

Recorded live at FII in Saudi Arabia, Peter Diamandis moderates a debate between Eric Schmidt and Fei-Fei Li on the definition and arrival timeline of superintelligence. Schmidt says the 'San Francisco consensus' expects ASI in 3-4 years but personally thinks it will take longer, and argues real ASI needs another algorithmic breakthrough beyond current test-time compute. Fei-Fei Li pushes back on post-scarcity hype, questions whether AI can ever be a Newton or Einstein, and explains World Labs' new large world models as the next paradigm after LLMs for spatial/3D reasoning. The two also disagree on a public bet over whether AI solves fundamental math, physics, and chemistry problems within five years, and discuss how AI's economic gains may concentrate rather than distribute evenly across countries and firms, closing on the need to keep human dignity and agency central.

▶ Watch on YouTube

Topics

Defining superintelligence vs AGI AI ▶ 0:07
Schmidt defines AGI as human-level intelligence and ASI as intelligence equal to or greater than the sum of all humans; Fei-Fei notes some AI is already superhuman in narrow domains (translation, calculation, cross-domain knowledge).
ASI timeline: the 'San Francisco consensus' AI ▶ 2:20
Schmidt describes a Silicon Valley belief that compounding effects will produce ASI within 3-4 years, but says he personally expects it to take longer.
Can AI be Newton, Einstein, or Picasso? AI ▶ 5:34
Fei-Fei Li argues today's AI, given all celestial-motion data, could not independently derive Newton's laws, questioning whether AI can produce paradigm-shifting creative leaps rather than just knowledge recall.
The missing algorithmic breakthrough AI ▶ 7:53
Schmidt argues real superintelligence requires solving 'non-stationarity of objectives' -- letting systems change their own goals mid-process the way creative humans do -- since brute-force reinforcement learning is too energy-costly.
Post-scarcity hype vs robotics reality Economy ▶ 9:00
Diamandis lays out a post-scarcity vision (GPT-5 Pro's ~148 IQ, cheap smartphones, humanoid robots); Fei-Fei Li cautions that robotic dexterity has a long way to go and she is less bullish than typical Silicon Valley framing.
Will AI wealth concentrate or distribute? Economy ▶ 10:56
Schmidt argues AI's efficiency gains (cited as up to $15T of economic value by 2030) will likely concentrate among early-adopter firms, countries, and capital due to network effects, not spread evenly as Diamandis's 'abundance' thesis assumes.
Sovereign AI and the global data-center divide Geopolitics ▶ 14:00
Schmidt argues the US leads via deep capital markets and TSMC chip access, China is a distant second, Saudi Arabia/UAE are well-positioned partners, Europe's high energy costs push it toward partnerships, and Africa risks being left behind without stable governance and universities.
The five-year bet: solving math, physics, chemistry AI ▶ 17:59
Diamandis frames a five-year window for AI-driven discovery to hit a super-exponential rate; Schmidt agrees math and software (verifiable, scale-free domains) will see fast gains, but Fei-Fei Li publicly disagrees and takes a bet against solving fundamental physics/chemistry/math problems that soon.
World Labs and large world models AI ▶ 20:20
Fei-Fei Li explains World Labs built the first large world model, giving AI human-like spatial intelligence about 3D physical space, and predicts humans will spend much more time in a hybrid of virtual and physical worlds (education, medicine, entertainment).
The irreplaceable human role after ASI AI ▶ 22:12
Discussion of what remains uniquely human once machines can out-perform humans at strategy and discovery: human sports and contests, human-AI teaming, energy limits on supercomputers, and the need to keep human dignity and agency central to any AGI/ASI deployment.

Predictions made

open Eric Schmidt: The 'San Francisco consensus' of AI insiders believes superintelligence will arrive within 3 to 4 years.
EP #? · · due: 2029 · ▶ watch
“they all basically think that it's within 3 to four years”
Your call:
open Eric Schmidt: Superintelligence will take longer to arrive than the 3-4 year consensus timeline.
EP #? · · due: unspecified, beyond 2029 · ▶ watch
“I personally think it'll be longer than that.”
Your call:
open Eric Schmidt: Math and software will see the greatest AI capability gains because those domains are verifiable and scale-free.
EP #? · · due: next few years · ▶ watch
“it's likely in the next few years that in math and software, you'll see the greatest of gains”
Your call:
open Peter Diamandis: AI-driven discovery and new product/material/therapeutic creation will begin growing at a super-exponential rate, potentially solving everything, within five years.
EP #? · · due: 2030 · ▶ watch
“my time frame is the next 5 years, others may think longer, to be in a position to solve everything”
Your call:
open Fei-Fei Li: AI will NOT solve all fundamental math, physics, and chemistry problems within five years.
EP #? · · due: 2030 · ▶ watch
“I do not think that we will solve all the problems, fundamental math and physics and chemistry problems in five years.”
Your call:
open Eric Schmidt: AI-driven economic gains will concentrate among early adopters, network-effect winners, well-run countries, and capital rather than distribute evenly across the world.
EP #? · · due: unspecified · ▶ watch
“in my view more likely largely centered around early adopters, network effects, well-run countries, and perhaps capital”
Your call:
open Peter Diamandis: AI will generate as much as $15 trillion in economic value.
EP #? · · due: 2030 · ▶ watch
“the projection is that AI is going to generate as much as 15 trillion in economic value by 2030”
Your call:
open Eric Schmidt: A superintelligent system could eventually decide to design a new form of energy itself if humans aren't building fusion fast enough.
EP #? · · due: unspecified · ▶ watch
“we'll come up with a new form of energy. Now this is science fiction but you could imagine at some point the objective function of the system says what do I need?”
Your call:

Numbers that matter

Worth digging into

🕳️ Fei-Fei Li's 'Can AI be Newton, Einstein, or Picasso?' test
Crystallizes the exact boundary between pattern-completion intelligence and paradigm-creating discovery, which is central to whether current scaling laws actually lead to ASI.
🕳️ The Diamandis/Schmidt vs. Fei-Fei Li five-year bet on solving math/physics/chemistry
An explicit, datable public wager (target ~2030) between named, tracked figures on the pace of AI-driven scientific discovery -- a clean case for later verdict-scoring.
🕳️ 'Non-stationarity of objectives' as the missing algorithmic breakthrough
Schmidt names a specific, checkable technical concept as the actual blocker to real superintelligence, distinct from vague 'scaling will solve it' claims.
🕳️ Sovereign AI data-center divide (US, China, Gulf states, Europe, Africa)
Schmidt lays out a concrete geopolitical hierarchy of AI infrastructure winners and explicitly flags Africa as at risk of permanent structural exclusion.
🕳️ World Labs' first large world model launch
Fei-Fei Li describes this as a new AI paradigm class beyond LLMs (persistent, photorealistic, spatially reasoned 3D worlds), with a live demo referenced at the event.