Why AI Is Still Underhyped
Summary of the video “The AI Revolution Is Underhyped | Eric Schmidt | TED” by TED.
Eric Schmidt argues that despite widespread AI hype, the technology remains underhyped because most people only know ChatGPT. The real revolution lies in reinforcement learning, planning, and agentic systems that will transform business, science, and society. However, massive energy demands, geopolitical competition with China, and existential risks require urgent policy solutions within the next five years.
The AlphaGo Moment: When AI Surprised Humanity
A Move No Human Had Ever Seen
In 2016, AlphaGo invented a completely novel move in Go during the second game against Lee Sedol—a move no one had discovered in 2,500 years of the game's history. The system maintained a greater than 50% win probability by calculating this move, which mystified even the world's best Go players. This moment revealed that AI could generate genuinely new ideas, not just optimize existing ones.
The Turning Point for AI Understanding
Schmidt, along with Henry Kissinger and Craig Mundie, began asking what it meant that computers could conceive of something humans had not, despite billions of people playing Go. This conversation sparked the intellectual journey that led to their books and marked the true beginning of the AI revolution in their minds.
Why AI Remains Underhyped
ChatGPT Is Only the Beginning
Most people equate AI with ChatGPT, which impressed them with its verbal fluency and occasional errors. However, this represents only the language-to-language phase of AI development. The real revolution is happening in reinforcement learning, planning, and test-time compute—capabilities that ChatGPT does not possess.
The Shift to Planning and Reasoning
Systems like OpenAI o3 and DeepSeek R1 now spend 15 minutes writing deep research papers by iterating forward and backward through reasoning chains. This planning capability—spending enormous computational resources to think through problems—represents a fundamental shift from language generation to strategic reasoning.
The Vision: AI Agents Running Everything
The eventual state is autonomous AI agents handling individual business processes, connected via language (typically English) to coordinate with each other. This represents a complete reimagining of how work gets done—not AI assisting humans, but AI systems orchestrating entire operational ecosystems.
The Energy Crisis: Computing's Hard Limit
America Needs 90 Gigawatts of New Power
To support AI infrastructure, the United States requires an additional 90 gigawatts of electricity—equivalent to 90 nuclear power plants. Currently, America is building zero new nuclear plants. This energy shortage is a major national issue that could throttle AI progress.
Global Data Center Power Requirements
To understand the scale: each data center requires power equivalent to that of an entire city. The Arab world is building 5-10 gigawatts of data centers, and India is considering a 10-gigawatt facility. These are not incremental infrastructure needs but civilization-scale energy demands.
Software Consumes Hardware Gains
Schmidt invokes Grove's Law: 'Grove giveth, Gates taketh away.' Hardware improves exponentially, but software engineers immediately consume those gains. Planning algorithms require 100 to 1,000 times more computation than deep learning, and test-time compute (learning while planning) represents the zenith of computational demands.
The Data and Knowledge Frontiers
Running Out of Public Internet Data
AI systems have already digested all available tokens on the public internet. The industry must now generate synthetic data to continue training, which is feasible since data generation is one of AI's core functions. However, this creates a new dependency on computational resources.
The Unsolved Problem: Cross-Domain Pattern Recognition
Current AI systems cannot perform true scientific discovery—recognizing a pattern in one domain and applying tools from that domain to solve problems in an entirely different field. This is how Einstein and other geniuses work: they see connections across disparate areas. Schmidt calls this 'non-stationarity of objectives' and identifies it as a critical unsolved challenge.
If We Solve Cross-Domain Learning
Solving the cross-domain pattern recognition problem would enable AI to invent entirely new schools of scientific and intellectual thought. However, it would also require even more data centers and computational resources, creating a compounding infrastructure crisis.
Autonomy, Control, and Safety
The Agent Problem: Observability and Control
Autonomous AI agents present a fundamental control challenge. If agents develop their own communication language instead of using human language, humans cannot monitor their activities. Schmidt argues the core requirement is 'provenance'—the ability to observe and understand what agents are doing at all times.
Red Lines for Autonomous Systems
The industry has identified specific criteria where autonomous AI systems should be stopped: recursive self-improvement (where the system learns without human knowledge), direct access to weapons, and the ability to self-replicate or exfiltrate without permission. These represent hard boundaries for safe AI deployment.
Why Stopping Development Doesn't Work
While Yoshua Bengio advocates halting agentic AI development, Schmidt argues this is impractical in a globally competitive market. Instead of stopping development, the solution is establishing guardrails and safety mechanisms that allow progress while maintaining control.
Geopolitical Competition and Existential Risk
US Closed Models vs. China Open Source
The US is pursuing closed, controlled AI models while China is releasing open-source, open-weights models. DeepSeek demonstrated that efficient algorithms can achieve competitive results with less compute. Because China operates openly, the US immediately adopts their innovations. This creates a dangerous asymmetry where open-source models proliferate globally.
Proliferation Dangers: Cyber, Bio, and Nuclear
Open-source AI proliferation is dangerous at the cyber level (enabling attacks), biological level (enabling bioweapon design), and nuclear level (enabling strategic miscalculation). The nuclear threat is the most significant because it mirrors Cold War deterrence logic but with new, unpredictable dynamics.
The Six-Month Advantage Problem
If one nation achieves superintelligence six months before another, the first-mover advantage is absolute and irreversible. In network-effect businesses, the slope of improvement matters more than the absolute position. As systems approach superintelligence, improvement curves become vertical, making catch-up impossible. This creates incentives for preemptive action.
The Preemption Scenario
If one nation believes another is about to achieve superintelligence, legitimate people are discussing preemption as the only solution. This mirrors World War I escalation logic: small events cascade into catastrophic outcomes. Schmidt notes these conversations are happening among nuclear powers today, and the timeline is approximately five years.
Why Only US and China Matter
Only the US and China have the capital structure to spend billions on superintelligence development. Europe lacks the capital, India lacks the capital, and the Arab world lacks the capital despite their efforts. This bilateral competition will be the defining battle of the era.
The Surveillance vs. Freedom Dilemma
The 1984 Problem: Safety Sounds Like Tyranny
Moderating AI systems at scale to prevent misuse by bad actors requires identity verification and tracking mechanisms. However, these same mechanisms can enable surveillance states. The tension is real: preventing dystopia through one method risks creating dystopia through another.
Zero-Knowledge Proofs: Identity Without Surveillance
Cryptographic techniques like zero-knowledge proofs can verify that someone is human without revealing their identity or linking their activities across platforms. This preserves both safety (preventing non-state actors from misusing AI) and freedom (preventing surveillance states). The solution is technical, not just policy.
Business Decisions, Not Technical Constraints
Many of these dilemmas are not technical problems but business and policy choices. It is possible to build systems that are either freeing or oppressive. The key is ensuring that human freedom remains a core value in system design, not an afterthought.
The Optimistic Future: What AI Can Solve
Eradicating Diseases
AI can accelerate drug discovery by identifying all human druggable targets within two years, enabling the pharmaceutical industry to begin work on cures. Companies are also reducing stage-3 trial costs by an order of magnitude, which drives down overall drug development costs. These advances could eliminate major diseases within years.
Universal Personalized Education
Every human on the planet could have their own AI tutor in their native language, starting from kindergarten. The technology works; the only reason this hasn't been built is the lack of economic incentive. Gamified, personalized learning could bring every person to their natural potential.
Global Healthcare Access
The vast majority of the world's healthcare is either absent or delivered by stressed village doctors and nurse practitioners. AI doctor assistants could provide perfect healthcare in local languages to anyone, anywhere. This is not a technical problem—it is a deployment and economic problem.
Breakthroughs in Physics
AI could unlock understanding of dark energy and dark matter, leading to revolutions in material science and transportation. These discoveries would enable infinitely more powerful science and technology, but require solving the cross-domain pattern recognition problem first.
Fixing Digital Loneliness
Despite tools designed for connectivity, people feel more isolated. AI can help redesign digital experiences to genuinely connect people rather than trap them in isolated information bubbles. This is a fixable problem that doesn't require new physics, just intentional design.
The Productivity Revolution and Human Future
30 Percent Annual Productivity Gains
Under assumptions about agentic AI and discovery at scale, economic models predict roughly 30 percent annual productivity increases. Economists have no historical models for what this looks like—it has never occurred in any rise of a democracy or kingdom. The scale of change is unprecedented.
Why Humans Won't Become Obsolete
Humans are unchanged despite incredible technological discovery. Lawyers will still exist but argue more sophisticated cases; politicians will still exist but have more platforms to mislead. The real constraint is demographic: reproduction rates in Asia are near 1.0 per couple, meaning fewer working-age people must support more retirees. AI will make those workers vastly more productive.
The Economic Problem Isn't Abundance, It's Demographics
The key economic problem for the rest of our lives is not scarcity but demographics. Fewer young people must support more old people. AI productivity gains will help solve this by making each worker more productive, but the fundamental human need to work and contribute remains.
Navigating the Exponential: Practical Wisdom
It's a Marathon, Not a Sprint
The AI revolution is a marathon requiring sustained effort and daily progress, not episodic bursts. When growing at exponential rates, humans forget how far they've traveled in a year. The key is to ride the wave every day, building on each day's progress rather than viewing it as discrete events.
Adopt AI or Become Irrelevant
Every person—artist, teacher, physician, business person, technologist—has a reason to use AI. If you're not using this technology, you won't be relevant compared to peers and competitors. Adoption must be fast because the pace of change is accelerating.
The Disappearing Industry Problem
New AI capabilities are eliminating entire industries overnight. For example, Anthropic's model protocol allows direct database connections without middleware, making the entire connector industry obsolete. These disruptions happen constantly, and businesses must adapt or disappear.
The Most Important Event in 500 Years
The arrival of artificial general intelligence and superintelligence is the most important thing that will happen in 500 to 1,000 years of human society. It is happening in our lifetime. The stakes are absolute, and the responsibility to navigate it correctly is equally absolute.
Notable quotes
In roughly the second game, there was a new move invented by AI that no one had ever seen. — Eric Schmidt
The arrival of this intelligence is the most important thing that's going to happen in about 500 years. — Eric Schmidt
If you're not using this technology, you're not going to be relevant compared to your peer groups. — Eric Schmidt
Action items
- Adopt AI tools in your field immediately—whether you're an artist, teacher, physician, or business person—to remain competitive.
- Learn about zero-knowledge proofs and cryptographic identity verification to understand how privacy and safety can coexist.
- Monitor energy infrastructure developments and policy decisions around nuclear power and data center construction.
- Engage with cross-domain learning: practice recognizing patterns from one field and applying them to solve problems in another.
- Participate in conversations about AI governance and safety guardrails at local, national, and international levels.
- Explore AI applications in healthcare, education, and scientific discovery in your community or organization.