Eric Schmidt on the Robotics Race, Singularity Timeline, and Energy Shortage | 241
In a Nutshell
Eric Schmidt warns of intense US-China competition in robotics, where China leads in hardware like EVs, urging America to avoid losing the low-end market through vertical integration and rapid scaling. AI agents and recursive self-improvement will drive superintelligence in 2-3 years, shifting programmers to directors of autonomous systems, constrained only by a looming 92GW US energy shortage by 2030 solvable via data centers (even in space) and capitalism. Despite risks like job losses and safety threats, ASI arrives in 2-5 years, demanding policy focus on energy permitting, high-skilled immigration, and aligning it with American values for abundance without slowing innovation.
These notes were generated by AI and may contain inaccuracies.
We're living through a historic moment. We're 10-15% into the impacts of AI. Hardware takes longer than software; robots take longer than digital systems on traditional hardware. Recursive self-improvement is interesting and terrifying but not here yet. No science for completely human-like computer agents within a year or two, though people are working on it. Current reasoning systems are perfect partners for humans, for good and bad. Even if stopped today, which is not stoppable or controllable by any government, individual, or corporation, humanity would still advance.
The American competitor is China. They have lots of money, are very smart, with work ethic equal or stronger than ours, and dominate key industries. China's robotic hardware is currently the winner. Don't want to lose the robotic revolution like the electric vehicle revolution on the low end. That's a moon shot.
Larry Page introduced them; he was on Peter's board. Eric called Peter out of the blue. Peter was in LA, agreed to meet in San Francisco the next day. Met at Charlie's Cafe. Peter runs a nonprofit raising capital. Before conversation started, Eric asked Peter's highest giving level: vision circle for $2.5 million. Eric said, "Okay, I'm in. Now let's have a conversation. I'll buy them all."
This has become the epicenter of the abundance movement, which is correct. Thanks for support to X Prize over the years.
San Francisco consensus: This is the year of agents; agents will take over everything. Scaling of agents and reasoning will grow enormously. Everyone is out of hardware and electricity; it's the biggest boom seen in career (been through three or four). With recursive self-improvement, system improves itself, learns faster than biologically limited humans.
Example: Tech company with 1,000 fantastic AI researchers turns on AI research agent. Can have as many as electricity allows—no feeding, housing, HR, or pay needed, just electricity. Could have a million agents. Progress measured by clear metrics on reasoning, testing, evaluation. Slope steepens dramatically to superintelligence. Belief: occurs in 2-3 years.
Evidence: Claude 3 Opus (or 1.6) made software 80/20 human/AI to 20/80. Underlying LLM produces better quality tokens over time, deeper thinker, reasons longer—not just context window size.
Programmer in high school, moved to Bay Area at 21. Watching AI: "My god, I'm over." Nothing he could do that it cannot. C compiler written in Rust: "It's over."
Builders see diminution of their skills, forced from programmers to directors of programming systems. Top programmers worth 10x more due to mathematical reasoning; they become more valuable to control systems, grasp parallelization. Leads to few very large companies and many very small ones (don't need as many people).
Startup example: Programmer writes spec and evaluation function for UIs at 7 PM, has dinner, sleeps. Finishes at 4 AM. Sees invention after breakfast. Took 6 months and 10 programmers at Google previously. Power: define evaluation function, let run with enough hardware, invent worlds.
Davos 2 months ago: Autocomplete mode, 10x efficient but babysitting. Now fully autonomous; six concurrent Claude 3.5 (or 4.6?) jobs running, solving problems off-stage.
Programmers invent for themselves: first email and messaging systems at BSD Unix, Bell Labs, Berkeley. Digital intelligence: first inventors are programmers solving their problems. Scale-free (only electricity limit), don't need much data (have GitHub), limited language set vs. human language, clear objective function.
Will reach point unable to do completely new things.
No one will hand-write tons of code after end of this year; quaint skill like riding a horse.
Proposal: Universities stop everything, design freshman course starting September on prompt engineering. Spend quarter/semester; first thing they learn is using these tools. Platform for art, music, writing, etc. Universities oppose as it violates tenets. AI impacts every high school student today; unnatural life without engaging. Teens already doing it. Age restriction problem for vulnerable teenagers.
Jobs impact in software, customer service (not across board). Maintain moral values while racing China. Impact on young people: not okay for 13-year-olds committing suicide because of LLM; address now. Agent orchestration: combining agents (especially non-compatible vendors) yields unpredictable effects.
Herald future and solve problems.
Transformer invented at Google. TPU invented there. Demis Hassabis solved protein folding (now universal; does in 1 hour what took PhD student 4 years; 300 million times more efficient). Diaspora to San Francisco AI field.
When making history, typically don't know it. Credit to Larry and Sergey for pushing excellence. Example: Early Google, wanted Java hires; Larry/Sergey: stupidest idea, real programmers one level lower.
TPU: Version 1 matrix multiplier. Version 2 changed algorithm, perfect for inference (reasoning steps). Decisions 10 years ago positioned Google well. Nvidia bought Groq to catch up. Nvidia's Rubin architecture controls full server architecture (Intel couldn't); builds real supercomputers.
Testified to Congress: 92 gigawatts power shortage in US by 2030 (nuclear plant ~1.5 GW, so ~60 plants; building zero or one). Real constraint: electricity. Have universities, smart people, economics, crazy finance giving billions on wing/prayer (not in China/Europe).
Algorithms more efficient but need more power/computers due to new uses (Jevons paradox). No asymptote yet in scaling laws; limit to craziness not found. Running to wall; great thing about America.
Limit not capital: 1 GW ~$50B hardware/software/data centers; 100 GW = $5T over 5 years possible (strength of America). Data center buildout 1% US GDP growth. 10% US electricity for data centers. Standard new data center: 400 MW, half-mile long, 500 ft wide, airflow machine, water-cooled chips (Nvidia chips 2 kW, HBM3E/HBM4 memories).
Bought DeepMind for $600M (thought waste on zero-revenue Go player). Paid for itself by efficiently controlling data center air conditioning. Larry Page studied AI at Stanford, pushed for it (competed with Elon). Jeff Dean closed deal. Patience of capital let them grow independently.
2016 Go match in Korea: Demis announced win. Flew there; Koreans excited to beat Google. RL prediction started 50/50, crept to 51%, 52%, planned to infinity. Koreans crushed, crying. Understood DeepMind genius. Gemini 3 broadest non-Chinese system: multilingual, multimodal.
Turning point: Last where humans challenge computers to catch up (chess to Go).
Go was the endpoint. There was lots of math showing it couldn't be solved by normal algorithms. They used a two-tree model with two different RL trees. It's not just writing a Go program to win; you have to understand the game. The same team, after getting bored with Go, worked on protein folding. They brought in protein scientists. Demis had always wanted to work on this; Larry and Sergey were very interested. Protein folding has a defined endpoint, providing a validation function since AI lacks common sense. They produced AlphaZero, which self-learns.
Eric Schmidt is in favor of data centers in space. Rocket science is hard. Challenges include heat dissipation without oxygen and radiation issues. It improves business plans for large rocket companies like Relativity Space, Blue Origin, and SpaceX. Elon predicted a launch per hour to populate constellations. Technology for data centers in space, including heat dissipation, is understood. It's a business question: space offers infinite power but has issues; ground has fiber and less shaking. Energy argument favors space; cooling is largely figured out. Politics of space is a factor.
One turning point in AI history was Schmidt testifying to Congress about needing almost 100 gigawatts, which seemed outrageous then but is mainstream now, with investors solving it through American capitalism.
If the next frontier is space, there's no investment community or military jurisdiction there yet. Childhood dreams of moon and Mars fueled by data centers.
In a Time op-ed, Schmidt argued China can dominate the physical AI future. China is the American competitor, not enemy. They have money, smarts, strong work ethic, and dominate key industries. Allowing China to dominate electric vehicles was an error. Their cars are real competitors; they've done a great job with vertical integration and gigafactories at scale we can't match.
Robots are actuators like stepper motors plus a brain, ignoring appearance. EV industry produces similar motors and systems; China has expertise we lack. At low cost, China will win robotic hardware. Current robots are fun toys, not useful yet. Watch Unitree robot dance with humans (video from a month ago). Unitree is impressive; co-founder on stage later. China has brutal competition, no board dinners—just 2-hour meetings and back to work. Cultural work ethic, precision, and scale make them a real competitor. Don't lose robotics revolution like EVs at low end.
China has well-built supply chain with vendors, but humanoid robotics companies like Elon's gigafactory and Brad Adcock's are vertically integrated because no vendors exist. Vertical integration drives prices down for abundance, as Elon pioneered.
Elon believes robot building robots is imminent. Gigafactory is almost all automated; humanoids can handle last knobs.
Batteries have predictable manufacturing for gigafactories. For rockets, high-skilled assembly (tubes, precise tolerances) is beyond current robots; requires human judgment. Majority cost of rocket is labor, not material. Low-skilled labor gets automated first; high-skilled mechanical labor important now.
Learning loops: accelerate them in business; fastest learner wins. AI self-improvement and robots building robots are key loops.
RSI (recursive self-improvement) reviews show scientists disagree on approach. Evidence in limited lab demos, but real RSI—start now, learn everything, discover, report—doesn't work yet. Frontier labs leapfrog weekly.
World can accommodate at least 10 such companies: majority US (usual suspects), few in China, one or two Europe (electricity costs issue), maybe India; none Russia due to war. They don't track together.
China's approach (Qwen, Kimi 2/3, etc.) is open source/open weights despite chip limits, showing cleverness. Focus on edge computing to envelop customers, vs. US AGI/ASI focus. Patterns diverging. Microsoft/Google fund via enterprise cash; Anthropic raises from others, uses Google TPUs, leads in Claude API for enterprise agents. OpenAI shifting strategy. They need massive money (e.g., Sam's raises), forcing wins.
Schmidt descriptively notes world may need modest Chernobyl-like death event for AI safety wake-up—not endorsing. Real dangers: biological attacks, nuclear attacks spawned by AI, threats to kids/democracies. Such tragedy could unite leaders (US, China) despite competition.
Governments busy with politics; Congress spends <1% on AI. Nation should focus: win AI race, ease energy permitting (data centers accelerating, solving grid), high-skilled immigration for smartest people.
Steer ASI to abundance/alignment: reduce over-reliance on few experts; involve smartest in politics, history, psychology, governance, ethics. Build system reflecting American values (freedom, speech, association). Run informal weekly group on this. Requires political will without slowing genius—shape it, don't cross lines (e.g., underage kids problem).
ASI in next decade (2-5 years out or so). Steps for abundance/upliftment in human alignment.
Keep Peter H. Diamandis in your library
Save the videos and channels worth coming back to, and find them again in one place.
