Public transcript

Eric Schmidt's Real Take on the AI Bubble

By: Catalaxis
Length: 19 min
Language: English
Text from: the video's captions
Transcribed: 14 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

This video compiles Eric Schmidt's remarks on AGI timelines, recursive self-improvement, and the AI investment bubble, followed by a speech from former OpenAI chief scientist Ilya Sutskever about AI's future impact. Schmidt argues super intelligence may arrive in six or seven years, that AGI requires systems able to set their own objective functions, and that current AI investment is more likely a new industrial structure than a bubble. Sutskever contends AI will eventually do everything humans can, because the brain is a biological computer, and urges people to pay attention and prepare.

Key points

Eric Schmidt rejects the three-year timeline for super intelligence, saying it will more likely take six or seven years.
General intelligence requires a system that can set its own objective function, and there is no evidence of that today.
The key unsolved technical problem is non-stationarity: the objective function keeps changing, as human goals do.
Humans are end to end while AI is middle to middle, so AI will be synergistic with humans rather than replacing them.
Schmidt expects systems that learn as you use them within six months, but full recursive self-improvement in 5 to 10 years.
He dates AGI at four to six years, with super intelligence — smarter than the sum of all humans — after that.
Despite executives admitting overbuilding, Schmidt believes the investment is a new industrial structure, not a bubble.
Ilya Sutskever argues AI will do everything humans can because the brain is a biological computer.

Questions it answers

01:00AGI timelines and the San Francisco narrative

When does Eric Schmidt expect super intelligence and AGI?

The San Francisco consensus says forms of super intelligence arrive in about three years; Schmidt disagrees and says maybe six or seven years. He notes savant-like systems in specific areas are not general intelligence, which requires the ability to set its own objective function.

Schmidt disagrees with the three-year timeline and says six or seven years.
A savant-level system is not yet general intelligence.
General intelligence requires setting its own objective function.

03:00The technical barrier to AGI

What technical problem stands between current AI and AGI?

The problem is non-stationarity: the objective function keeps changing, as human goals do daily, and no LLM algorithm handles that yet. Schmidt suggests the test is whether a system could derive Einstein's special and general relativity using only information available in 1902, and most people believe this will be solved through analogy.

The key unsolved problem is non-stationarity of objective functions.
The test: could AI derive Einstein's relativity from 1902 information?
Most people believe the solution will come through analogy.

03:50Humans are end to end, AI is middle to middle

Why won't AI simply replace humans in the near term?

Humans are end to end with complete understanding and easy pivoting, while AI is middle to middle: it must be prompted, its answers validated, and iterated with more context. Schmidt says that for the next few years each of us will have assistants that are incredibly helpful on our command.

AI must be prompted, validated, and iterated.
Humans define the objective function and prompt.
AI will be synergistic with humans rather than replacing them.

05:55Recursive self-improvement and super intelligence

What is recursive self-improvement and when does it arrive?

Recursive self-improvement is where the system learns itself, and Schmidt expects systems that learn as you use them within six months, though he estimates full recursive self-improvement is 5 to 10 years away. He dates AGI at four to six years, and super intelligence, where the system is smarter than the sum of all humans, after that.

No evidence of recursive self-improvement yet.
Systems that learn as you use them expected within six months.
Super intelligence test: a proof no human can understand but is known true.

08:15National security and race conditions

How does AI affect national security and geopolitics?

Schmidt describes a scenario where a country with a million AI researchers, needing only electricity, achieves steep innovation slopes in a network-effect business. The opponent may conclude it can never catch up, creating a race condition of preemption, and every aspect of human experience including democracy will be touched.

A million AI researchers need only electricity.
A leading country may believe the opponent can never catch up.
This creates a race condition of preemption.

09:45Is the AI investment boom a bubble?

Does Eric Schmidt think AI investment is a bubble?

Executives say there is overbuilding and over capacity in two or three years, but each believes they'll be fine, which is a classic bubble. However, Schmidt has never seen hardware capacity not taken up by software, so he thinks it's more likely a whole new industrial structure, and calls the discovery of LLMs and foundation models historically important.

Executives admit overbuilding but say they'll be fine.
Reasoning models are thousands of times more expensive than a Google search.
LLMs are a historically important discovery credited to OpenAI.

12:50Ilya Sutskever on AI's extreme future

Why does Ilya Sutskever believe AI will do everything humans can?

He argues that all of us have a brain and the brain is a biological computer, so a digital computer should be able to do the same things. He urges people to use AI and watch what the best systems can do, because seeing with our own eyes will generate the energy needed to face the greatest challenge and reward of humanity.

The brain is a biological computer, so a digital one can match it.
AI progress will make the rate of progress extremely fast.
The challenge of AI is the greatest of humanity, with the greatest reward.

Notes

AGI Timelines

The "San Francisco narrative" holds that agents, recursive self-improvement, and scale-free problems like programming and math will produce forms of super intelligence in about three years.
Eric Schmidt disagrees with three years but agrees it may happen in six or seven years.
Savant-level systems (a chemist, physics, or mathematician savant) are not general intelligence. General intelligence means the system can set its own objective function, and there is no evidence of that today.

The Technical Barrier

Schmidt is writing a paper on the core problem: non-stationarity of mathematical proofs, where the objective function keeps changing the way human goals do daily.
The test: using only information available in 1902, can a system derive what Einstein did with special and then general relativity? We cannot do that today.
Most people believe the solution will come through analogy — geniuses understand one area deeply and apply ideas to a completely different domain. If that is solved, Schmidt says "it's over" and we get AGI.

Humans vs. AI Today

Humans are end to end; AI is middle to middle. AI must be prompted, validated, and iterated with more context before producing business value.
For the next few years, AI will be synergistic with humans: we define the objective function, do the prompting, and AI does the work in the middle.

Recursive Self-Improvement and Super Intelligence

No evidence of recursive self-improvement yet; Schmidt has funded startups claiming to be close, which suggests 5 to 10 years.
Systems that learn as you use them should appear within six months.
Reasoning models are many thousands of times more expensive than a Google search, which explains the massive hardware investments.
Schmidt's dates for AGI: four to six years. Super intelligence is smarter than the sum of all humans; its test is proving something true that no human can understand.

Bubble or New Industry?

Executives admit overbuilding and over capacity in two or three years, but each says "I'll be fine and the other guys will lose all their money" — a classic bubble.
Schmidt has never seen hardware capacity not taken up by software ("Grove giveth and Gates taketh away"), so he believes this is more likely a whole new industrial structure.
He credits the team at OpenAI for the historically important discovery of LLMs and foundation models.

Ilya Sutskever's Warning

AI will eventually do all things humans can do, because the brain is a biological computer and a digital brain can do the same.
He urges people to use AI and watch what it can do, since seeing with our own eyes generates the energy to face what he calls the greatest challenge — and greatest reward — of humanity.

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