The State of AI: Humanoid Robots, AI Copyright Wars & China’s Growing Influence w/ Salim Ismail #157
Summary
Figure AI is turning humanoids from a prototype story into a manufacturing and unit-economics story. Peter says its new factory can produce 12,000 robots annually and scale to 100,000, while Brett Adcock sees demand for 100,000 today and volume pricing below $20,000. Peter reverses his prior skepticism: when one machine learns a task, “instantly five million other robots know how to do that task.”
The China contest mixes legitimate security exposure with incumbent self-protection. Salim believes both that DeepSeek is state controlled and that OpenAI wants government help blocking a competitor, calling Sam Altman’s IP complaint “a little rich.” Peter argues the larger threat is “the rogue state” or individual misusing advanced AI and proposes a Bretton Woods- or Asilomar-style international safety framework.
Copyright access may be essential to the AI race, but uncompensated extraction is not a durable business model. Salim says copyright has been “misused for the last 50, 80 years,” yet creators still need payment as models run out of data. Peter backs ProRata.ai, which estimates each source’s contribution to an answer and divides revenue accordingly—a potential royalty layer for an industry already absorbing roughly $1 billion a day.
AI is beginning to reprice medical expertise through discovery, diagnosis and personalized guidance. Peter highlights a Stanford report on an AI-identified peptide reportedly producing Ozempic-like weight loss without its side effects, plus a case in which Claude 3.7 from Anthropic—Peter thinks that was the model—identified multiple myeloma from three-month-old records after physicians missed it. Salim claims doctors give the wrong diagnosis about 30% of the time; Peter’s practical call is to obtain your data and seek AI-assisted “second and third diagnoses.”
Education systems that ban AI are preserving old assignments while other countries raise student capability. Estonia and Beijing are introducing AI education, while all 12 teenagers Salim surveyed at the summit said their schools prohibited classroom use. Peter’s alternative is to give middle-schoolers genuinely hard problems and AI tools; Salim calls the coming divide “a reckoning” in which non-users are “left behind very fast.”
The next investable stack reaches from machine-readable content to physical agents and neural interfaces. Google’s Gemini robotics work aims to connect general AI with different machines, Ilya Sutskever’s superintelligence startup reportedly discussed raising $2 billion at a $30 billion valuation, and Science is developing neural-stem-cell interfaces that grow into the brain “like roots into the soil.” Meanwhile, an OpenAI employee’s call is that “99.9% of web attention is about to be LLM attention.”
Both speakers expect turbulence before abundance, but reject dystopia as predetermined. Mo Gawdat forecasts 5–12 years of dystopian risk; Salim broadens that to “a scary decade or two,” including possible nuclear proliferation. Peter frames the choice as “Star Trek versus Mad Max”: two futures held “in superposition,” with the outcome shaped by what people believe and how they act now.
Deep dive
1. Humanoids cross from prototype speed into factory scale
Peter’s production snapshot: Figure AI’s new facility can make 12,000 robots annually and scale to 100,000. Figure 02 is in production, Figure 03 is nearing announcement, and the company iterates a new design every 12–18 months after reaching delivery just 31 months from founding.
Brett Adcock told Peter that customers could absorb 100,000 units today. Figure is already at BMW, has signed a large logistics customer and expects pricing to fall from probably below $100,000 toward below $20,000 at volume, with entry into homes coming “very shortly.”
Peter’s explicit reversal matters: “I need to eat crow and revise my assessments.” He had doubted humanoid form factors and timelines, but now argues that shared learning changes the calculus—one robot can master a task and transfer it across a fleet, making the system “unbeatable over time.”
2. AI rivalry requires security rules and a workable copyright market
Asked whether DeepSeek is state controlled and whether OpenAI is using Washington as a competitive blocker, Salim answers that “both are very true.” He believes China uses DeepSeek as it used TikTok, while calling Sam Altman’s concern about IP violations “a little rich” given how OpenAI obtained much of its training data.
Peter relays Alvin Wang Graylin’s framing: “the greatest foe that we have is not China,” but a rogue state or individual using advanced AI to harm others. Peter favors collaboration among China, Europe, the Middle East and the U.S. to identify those actors.
Peter proposes combining the logic of Bretton Woods with the Asilomar guidelines: a global convening that establishes clearer AI safety principles. The reason is systemic rather than diplomatic—“one area goes rogue, it’ll affect the whole world.”
On copyright, Salim favors access because protection has been “misused for the last 50, 80 years,” particularly through Disney-driven extensions, but insists owners need compensation. Peter—disclosing that his fund invested—offers Bill Gross’s ProRata.ai: estimate each source’s share of an output and apportion money accordingly. With roughly $1 billion a day entering AI, incentive alignment becomes increasingly consequential.
3. AI is exposing the limits of human-only medicine
Peter highlights Stanford Medical School work that used AI to identify a peptide reportedly producing the same kind of weight loss as Ozempic without its side effects. His broader claim is that models will detect biological relationships humans cannot see, affecting cancer and eventually “all diseases.”
His sharpest diagnostic example involves a summit faculty member whose relocation after a house fire led a new doctor to discover multiple myeloma. When the patient put medical data from three months earlier into a model, Peter thinks he was using Claude 3.7 from Anthropic; it “instantly pegged it” after physicians had failed to diagnose the condition.
Salim says patients receive the wrong diagnosis about 30% of the time, while recalling Daniel Kraft’s point that several hundred cancer papers may appear daily—far beyond any oncologist’s reading capacity. Peter mentions a startup aiming to release a fully fledged, free AI doctor by year-end. His recommendation is to obtain one’s medical data and use models for additional opinions.
Fountain Life illustrates the personalized endpoint: 200 gigabytes covering the genome, full-body MRI and other knowable data, queryable through an AI system. Peter’s concrete use case is photographing a restaurant menu and asking what to eat given current blood markers and genetics.
4. Peter argues that AI education should make assignments harder
Estonia plans ChatGPT access across secondary schools, while Beijing is introducing AI courses in primary and secondary education. Salim says the reported programs require students to enroll in at least eight hours of AI during the year, then assumes that means per week; by contrast, all 12 teenagers he surveyed said their schools told them not to use AI.
Salim calls today’s institutional resistance an “immune-system response” from legacy education actors. His prediction is “a reckoning” arriving faster than expected: schools and students that do not integrate AI throughout education will be “left behind very fast.”
Peter concedes that conventional writing and mathematics exercises become trivial with AI. His solution is not to withhold the tools but to raise the target—ask middle-schoolers to design a new socioeconomic system or build a difficult program, then let AI expand what they believe they can attempt.
Salim’s benchmark: reconstructing Einstein’s relativity formalism took him a university year; with AI answering successive “why is this relevant?” questions, he thinks he could relearn it in roughly three days. Teachers should become “the guide on the side” in flipped classrooms, though he expects unions to resist until they are forced to adapt.
5. Embodiment turns AI from an information engine into an actor
Peter separates robotics into hardware and the intelligence controlling it. Google’s Gemini robotics work points toward a general software layer able to interface with many physical systems—the “real Holy Grail” because movement is where code finally affects logistics, health care, finance and the material world.
Salim links embodiment to theories that consciousness emerges through adaptation to physical surroundings. He is equally struck that Google described no plans for commerce, applauding research and open-model work that “just uplifts everybody.”
Ilya Sutskever’s superintelligence startup was reportedly discussing a $2 billion funding round at a $30 billion valuation. Peter says superintelligence is intended to mean something far beyond coworker-like AGI: a capability able to solve “incomprehensibly hard problems,” but whose immense power also creates immense navigation risk.
Salim’s pushback—worth keeping: “What the hell do we mean by superintelligence?” Peter’s provisional definition is AGI matching the best human across every field, followed by ten doublings—1,000 times more capable—to reach digital superintelligence that could tackle wave-particle duality, unified physics and civilizational governance. Salim adds examples including dark matter and doubling the human lifespan.
6. AI-native markets reward participation, youth and machine-readable infrastructure
An OpenAI employee’s provocation is that web content is still written for humans even though “99.9% of web attention is about to be LLM attention.” Salim sees a new infrastructure layer in which agents navigate pages and return only what is relevant to each user.
Cathie Wood’s summit call emphasized humanoids and robotaxis, with Tesla’s Cybercab expected to be up and operating in Texas that summer, followed by Los Angeles and, she hoped, nationwide deployment.
Peter relays investor Dave Blundin’s rankings: MIT first for AI and unicorn formation, USC second and Stanford third. The founder-age bell curve has shifted from the 30s to roughly 20–23, while an estimated 75% of incoming MIT freshmen want to start a company before graduating.
Two companies out of Exponential Ventures and MIT went from zero to $2 billion valuations in two years, Peter says. The strategic lesson from the investing panel was participation: once a technology becomes mission-critical, capital provides a way to keep learning because “your mind goes where your money goes.”
7. Interfaces are moving from implanted wires to living tissue and AI twins
Max Hodak’s Science Corporation is pursuing a biohybrid neural interface rather than inserting 1,000–2,000 fine filaments into the neocortex. Neural stem cells sit in millions of nutrient-filled microwells, connect to electronics, then grow axons and dendrites into the brain “like roots into the soil.” Peter says this avoids the neuronal destruction he attributes to filament insertion.
The design had been demonstrated in animal models, with primates planned later that year and humans hoped for a couple of years afterward. Against present communication rates of roughly 10–40 bits per second, it made Ray Kurzweil’s prediction of high-bandwidth brain-computer interfaces by 2033 newly plausible to Peter.
Salim’s underlying point is epistemic humility: the only neuroscience definition he could once get broad agreement on was a “fractal, chaotic information-processing system that may or may not generate consciousness.” Effective interfacing may matter more than full understanding because the brain itself can optimize the new connections.
HeyGen’s two-minute capture demonstrated another interface: high-fidelity avatars able to speak 17 languages. Salim wants an AI twin that remembers decades of work and becomes “so much smarter than the real me”; he also observes that younger NFT communities already use Ethereum or Bitcoin, not dollars, as their intuitive unit of account.
8. The path to abundance runs through a volatile institutional transition
Mo Gawdat expects eventual abundance but warns of 5–12 years of dystopian danger first. Salim agrees that “it’s always darkest before the dawn,” extending the concern to a decade or two. He cites a geopolitical analyst he follows who says the U.S. pulling away could lead every country to seek nuclear weapons defensively; he calls that a bad outcome. Drones at least reduce direct human participation in warfare.
Peter’s framing is “Star Trek versus Mad Max”: society holds two AI futures “in superposition,” one threatening and one profoundly stabilizing. Salim’s basis for optimism is that technology may be humanity’s primary driver of progress and that greater intelligence “tends to become wiser and more benevolent,” though the transition remains highly volatile.
Google’s Wing shows the practical upside: drone delivery operating in Dallas, with potential emergency uses such as moving a defibrillator to a cardiac patient. Starline similarly creates three-dimensional telepresence that feels physically present; Peter says Google plans commercialization by the end of the year.
Palmer Luckey’s closing lesson was to find where one can create the most impact and “shut up and go do that.” He sold Oculus for roughly $2.2–$2.3 billion not principally for the payout, but because Mark Zuckerberg promised $1 billion annually for ten years of development—an R&D commitment Luckey could not match independently. Meta ultimately spent about $60 billion.