Eric Schmidt: Everyone Is Wrong About AI. Here's What's Actually Coming.
- Length
- 5 min
- Language
- English
- Text from
- the video's captions
- Transcribed
- 14 October 2026
Summary
Eric Schmidt argues AI is underhyped because progress has moved beyond chatbots like ChatGPT into reinforcement learning, planning, and test time compute. He sees the eventual state as agents running all business processes and communicating in language with each other. The main constraints he identifies are energy and hardware, with America needing another 90 gigawatts of power, plus data scarcity and the unsolved question of whether AI can invent genuinely new knowledge.
Key points
- Schmidt argues AI is underhyped because most people still judge it by ChatGPT, while reinforcement learning has since enabled planning and strategy.
- He cites OpenAI 03 and DeepSeek R1 as examples of systems that plan by going forward and back, and deep research systems that spend 15 minutes of supercomputer time writing papers.
- He predicts the eventual state is agents running all business processes, concatenated together and speaking English among themselves.
- The first major limit is energy: America needs another 90 gigawatts of power, equivalent to 90 nuclear power plants, and is building zero.
- The Arab world is building 5 to 10 gigawatt data centers and India is considering a 10 gigawatt data center; each data center needs power on the scale of a city.
- Planning may require a factor of a hundred to a thousand more computation, and test time compute — planning while learning — is the zenith of computation needs.
- The second limit is data: we have run out of it, but generating data is something these systems can do.
- The third limit is knowledge itself: current systems cannot invent something completely new the way Einstein did by applying patterns from one field to another, a problem Schmidt is working on under the term non-stationarity of objectives.
Questions it answers
00:00Why AI is underhyped
Why does Eric Schmidt think AI is underhyped?
Most people judge AI by ChatGPT's verbal brilliance, but since then reinforcement learning has enabled planning and strategy, as seen in OpenAI 03 and DeepSeek R1. Schmidt says the eventual state is agents running all business processes, concatenated together and speaking English among themselves.
- Reinforcement learning now enables planning, as seen in OpenAI 03 and DeepSeek R1.
- Deep research systems spend 15 minutes of supercomputer time writing papers.
- Schmidt sees agents eventually running all business processes and speaking English to each other.
02:01Energy as the real limit
What are the real limits on AI scale, and how serious is the energy problem?
Schmidt says the real limit is energy: America needs another 90 gigawatts of power, equivalent to 90 nuclear power plants, and is building zero. He points to the Arab world building 5 to 10 gigawatt data centers and India considering a 10 gigawatt data center, noting each data center needs power on the scale of a city.
- America needs another 90 gigawatts of power, equal to 90 nuclear plants.
- The Arab world is building 5 to 10 gigawatt data centers; India is considering 10 gigawatts.
- Planning may require a factor of a hundred to a thousand more computation.
03:01Compute demand keeps growing
Why won't algorithmic improvements reduce the need for power and compute?
Schmidt invokes the old rule 'Grove giveth, Gates taketh away' — hardware gains get consumed by software. Planning with today's algorithms requires at least a factor of a hundred, maybe a thousand, more computation, and test time compute, which combines planning with learning during planning, is the zenith of computation needs.
- Software always consumes hardware gains: 'Grove giveth, Gates taketh away.'
- Test time compute combines planning with learning during planning.
- Data scarcity can be solved by generating data.
04:24The limit of knowledge
What is the third limit Schmidt identifies, and can AI invent truly new knowledge?
Schmidt asks what the limit of knowledge is: today's systems cannot do what brilliant humans like Einstein did, spotting a pattern in one area and applying its tools to a completely different one. He is working on this problem, known technically as non-stationarity of objectives, and if solved it would require even more data centers and could enable entirely new schools of scientific and intellectual thought.
- Current systems cannot transfer patterns across unrelated fields the way Einstein did.
- The technical term for the problem is non-stationarity of objectives.
- Solving it would require even more data centers and could enable new schools of thought.
Notes
Why AI is underhyped
- Most people's mental model of AI is ChatGPT — brilliant, verbal, but mistake-prone. That was the
Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.