Public transcript

Google前CEO施密特：AI競賽已經變成國家權力戰，未來10年最危險的變局來了

By: New SciTech 新科技
Length: 1 h 2 min
Language: English
Text from: the video's captions
Transcribed: 14 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

In this Princeton conversation, former Google CEO Eric Schmidt traces AI's rise from symbolic reasoning's failure through deep learning, AlphaGo and the transformer paper, to today's models that can prove theorems and rival graduate students. He covers the data center build-out and its power demands, argues China is within 6 months of the US and will dominate with free open-weight models, and warns that AI is offense-dominant in cyber and that data centers are military targets. He closes with views on white-collar job displacement, robotics, education, and audience questions on AGI, consciousness and finance.

Key points

Schmidt describes an OpenAI-generated proof of an 80-year-old theorem, produced by a model not trained on math, as possibly one of the most historic moments in human history.
He traces AI's rise from symbolic reasoning's failure, through the AI winter, to deep learning math in the 1980s, Google's 2011 cat discovery on YouTube, DeepMind's 2015 AlphaGo win, and Google's 2017 paper 'Attention Is All You Need'.
He argues computers now compete with top graduate students in basic sciences, writing and analysis, but lack unwritten human knowledge and wisdom, and he is funding research into the limits of this 'analytical superintelligence without wisdom'.
Data center build-out has added 1% to US GDP growth; 1 gigawatt of power costs about 50 billion dollars, and the industry needs 90 to 100 gigawatts by 2030, with new data centers built 'behind the meter' off the grid.
He believes chip export restrictions gave the US less advantage than expected, estimating China is within 6 months of the US based on Deep Seek V4 evaluations.
China's top models are open weight and open source, and because they are free, Schmidt predicts the majority of the world's models will be Chinese.
He views AI as offense-dominant in cyber, citing Anthropic's Methos cyber attack model released only to about 40 companies and the US government.
AI is currently displacing low-level customer service and entry-level software jobs, with mass white-collar effects expected within 5 years, and a possible robotic displacement of hands-on jobs in about 10 years.
He rejects the P(doom) polling process, does not believe LLM next-word prediction alone will reach AGI, and does not think AI will become conscious because it is not biological.
He recommends a mandatory freshman 'LLM test' at every university, analogous to old TB and swimming tests, so students learn to work with models that will mediate their lives.

Questions it answers

00:00Opening and the breakthrough moment

Why does Schmidt open by thanking Princeton, and what recent event does he highlight?

He credits Princeton as the binary pivot point of his success, then describes OpenAI's computer-written proof of an 80-year-old theorem by a model not trained on math, which he calls possibly one of the most historic moments in human history.

Princeton begat Berkeley, family, money and opportunities for him
The proof came from a model not trained on math
He views LLMs as synthesizing knowledge into vector spaces humans cannot scan

03:49From AI winter to deep learning

How did we get from failed symbolic reasoning to today's AI?

Symbolic reasoning failed, causing the AI winter; deep learning math emerged in the 1980s, Google discovered the concept of a cat on YouTube in 2011, AlphaGo won in 2015, and the 2017 paper 'Attention Is All You Need' spawned today's competition.

Deep learning began as pure mathematics
AlphaGo was thought incomputable
One 2017 paper created the modern industry

06:53Intelligence without wisdom

What can these systems do, and where do they fall short?

Computers now rival top graduate students in basic sciences, writing and analysis, but miss unwritten human knowledge and wisdom; he is funding research into whether this is 'analytical superintelligence without wisdom'.

They are not human in their intelligence
Thousands of unwritten facts elude them
No one has solved training data for wisdom

10:30Data centers and power

What is the scale of the data center build-out and what limits it?

The build-out added 1% of GDP growth; 1 gigawatt costs about 50 billion dollars, the industry needs 90–100 gigawatts by 2030, and new centers are built off the grid because interconnection queues are full for 7 years; the limit is financial, not physical.

An average nuclear reactor is 1 gigawatt
New data centers are 'behind the meter'
America's empty land and financial system enable the build

16:58Regulation and outrage

What lessons does Schmidt draw from regulating the internet and social media?

He says the internet's hands-off approach worked, but social media's did not, because maximizing revenue means maximizing outrage; he distinguishes free speech for humans from 'boosted speech' amplified by bullhorns.

Outrage-driven engagement misaligns business and society
He supports human free speech but limits on amplification
He expects AI to be regulated in different ways for different reasons

20:00China and open source

How close is China, and why does open source matter?

Chip restrictions were meant to give a 5–10 year advantage, but Deep Seek V4 suggests China is within 6 months; China's open-weight free models mean most of the world's models will likely be Chinese, and he warns the US is not meeting the competition.

China imbues AI in everything with state support
Free Chinese models attract startups and researchers
Only large countries like the US and China have the required scale

29:00Superintelligence and security

How does Schmidt connect AI to national security and cyber attack?

Data centers become military targets, and he proposes deterrent strategies analogous to nuclear mutual assured destruction, which has never been tested; AI cyber tools are offense dominant, as shown by Anthropic's Methos released only to about 40 companies and the US government.

Superintelligence may arrive within 10 years
He argues for strong offense as well as defense
MAD has never actually been tested

34:14Jobs and robotics

Which jobs does AI threaten, and on what timeline?

AI is displacing low-level customer service and entry-level software jobs now, with mass white-collar effects within 5 years; in about 10 years robots like Unitree's $5,000 R1 may displace hands-on jobs, which he sees as the real concern.

Two-thirds of US jobs are hands-on and safe for now
A Cloudflare-style 20% layoff shows category-based cuts
The US is bad at managing worker transitions

Notes

AI's arrival and why now

Eric Schmidt opens by crediting Princeton as the binary pivot point of his life: Princeton begat Berkeley, family, money and opportunities.
He calls OpenAI's computer-written proof of an 80-year-old theorem possibly one of the most historic moments in human history, noting the model was not trained on math.
He explains LLMs as systems that synthesize knowledge into a small vector representation space and collect facts across it in ways humans cannot; many believe this is the beginning of artificial superintelligence.

History of the field

Symbolic reasoning (a 'Spinoza model of knowledge') failed, causing the AI winter of the '80s and '90s.
In the 1980s, mathematicians at the Institute for Advanced Studies proved the theorem behind today's deep learning; models were built in the early 2000s.
In 2011, Google applied it to YouTube and the system discovered the concept of a cat from uploaded cat pictures.
Key milestones: DeepMind's 2015 AlphaGo win via reinforcement learning, and Google's 2017 paper 'Attention Is All You Need', which spawned today's industry.

Limits: wisdom and unwritten knowledge

Computers now compete with top graduate students in basic sciences, writing and analysis, perhaps art.
They miss thousands of unwritten pieces of knowledge (his 'joint ventures don't work, rev shares don't work' example) and lack wisdom; no one has solved getting 'training data for wisdom'.
In 15 years, people in his industry believe essentially all human diseases could be solved, plus new energy, materials and chemistry; a thousand years from now this period will be recorded as an epochal change, like the Enlightenment.

Data centers and power

The build-out added 1% of GDP growth to America; 1 gigawatt costs about 50 billion dollars; the industry needs 90–100 gigawatts by 2030 (an average nuclear reactor is 1 gigawatt).
New data centers are 'behind the meter', off the grid, because interconnection queues are full for 7 years; the limit is financial, not physical, and land plus the American financial system enables it.

China, competition and security

Chip export restrictions (Trump 1, Biden, Trump 2) were meant to give a 5–10 year advantage; Deep Seek V4 suggests China is within 6 months.
China's models are open weight and open source, and being free, most of the world's models will likely be Chinese.
Data centers become military targets; he proposes deterrent strategies analogous to nuclear mutual assured destruction, which has never been tested.
AI cyber tools are offense dominant; Anthropic's Methos was released to only ~40 companies and the US government.

Jobs, growth and Q&A

AI currently displaces low-level customer service and entry-level software jobs; mass white-collar effects within 5 years; robotics (e.g. Unitree's $5,000 R1) may displace hands-on jobs in ~10 years.
He rejects P(doom) polling, doubts LLMs alone reach AGI, and says AI will not be conscious because it is not biological.
Policy: a mandatory freshman 'LLM test' at every university; most financial investing will become automated.

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.