Balaji on the State of AI/AGI, Bitcoin & America’s Incoming Collapse w/ Dave & Salim | EP #191
Summary
- Balaji Srinivasan rejects the “one digital god” thesis: frontier models are leapfrogging rather than separating by orders of magnitude, while espionage, imitation, and open Chinese models keep capability distributed. Dave Blundin takes the opposite side, arguing recursive self-improvement has begun and could reveal a dominant ASI within six months, with two years his outer bound. Peter Diamandis argues that government intervention could preserve a “polytheistic” market even if the natural dynamic is winner-take-all. The unresolved trade is whether algorithmic gains and inference-time compute outrun capital scale—or whether one system consolidates.
- Today’s AI is primarily amplified intelligence: its bottlenecks are the human who points it and the expert who verifies what comes back. Balaji calls prompting a high-dimensional vector and says “everybody’s a CEO, everybody’s a manager,” but garbage prompts still yield polished “AI slop.” In practice, “AI doesn’t take your job. It takes the job of the previous AI,” creating recurring spend on better models, internal AI teams, process optimization, proctoring, and verification.
- Balaji’s macro map has two rising powers—China controls the physical stack while the free internet controls the digital stack—and the Western postwar system loses leverage between them. China’s advantages are manufacturing, automation, supply chains, and sovereignty; the internet’s are AI, crypto, free speech, free markets, and decentralization. His prescription is not to imitate Beijing but to become “more America than America,” because the open internet offers capabilities China cannot copy without weakening its control system.
- Bitcoin is “high-voltage money,” not necessarily the rail for every small payment, while wrapped and off-chain BTC can support internet commerce and machine payments. Balaji declined to disclose his Bitcoin percentage but said, “Only hold in USD what you can afford to lose.” Separately, Peter said that if fiat were unavoidable, he would prefer SGD or AED over Western currencies. Balaji’s portfolio mechanism is the “billionaire flippening”: somewhere between $100,000 and $1 million per BTC, half the world’s billionaires become crypto-rich, and another 10x sharply dilutes fiat wealth, banks, and states.
- Crypto’s machine advantage begins with programmability: an AI or robot can receive unlimited wallets and private keys in fractions of a second, then transact at scales, speeds, and complexity unavailable to a bank account. Balaji therefore expects crypto exchanges to become the new cross-border banks, though Blundin pushes back that lending and credit underwriting survive the migration of deposits. By roughly 2035, Balaji expects cryptocurrency to become a test of sovereignty: “You’re not a country if you don’t have cryptocurrency.”
- AI may accelerate medicine dramatically, but Balaji disputes forecasts that ignore the state’s control over trials, prescriptions, reimbursement, and professional licensing. He separates AI’s strength at synthesizing papers, scans, and known biological relationships from discovering points “outside the envelope,” while identifying regulation and risk tolerance as the binding constraints. Bitcoin and longevity share his deeper refusal of managed decay: traditional finance asks people to lose a little wealth annually, and traditional medicine accepts losing a little health.
- Network states are Balaji’s hedge against institutional decline: globally recruited “dark talent,” online capital, physical community, and more permissive jurisdictions combine into startup societies. His immediate playbook is to go direct, learn AI and crypto deeply, move toward Texas or Florida if remaining in the US, take coins off exchanges, build trusted local tech communities, and travel through rising non-Western hubs. The risk behind that advice is political: he expects “red and blue against tech” as automation, data centers, biotech, and visible founder wealth become shared scapegoats before any abundance arrives.
Deep dive
1. The internet, not AI, is the upstream force remaking 2035
Asked whether 2035 would be recognizable, Balaji answered that the continents would remain but much of the societal superstructure would not: “Very few institutions that predated the internet will survive the internet.” He treats the network as upstream of AI, crypto, robotics, China’s rise, and even contemporary political organization.
His evidence is political as much as technological. Balaji claimed that Twitter elected Donald Trump, then deplatformed him, and X later reelected him; he also said Twitter caused Brexit and the political fracas seen elsewhere. His point is categorical: the internet is already reorganizing legitimacy, yet remains underestimated because it is constantly “in front of your face.”
Peter Diamandis supplied Ray Kurzweil’s benchmark: the change from now to 2035 could equal the progress between 1925 and 2025. Balaji agreed, contrasting the relative stability of roughly 1980–2015 with the rapid institutional movement of the decade since.
The investor-relevant split in Balaji’s model is stark: China is “everything physical,” while the internet is “everything digital.” China took manufacturing and military supply capacity from red America; AI and crypto disrupted blue America’s control of media and money.
2. China and the free internet become rival operating systems
Balaji uses “China” as shorthand for land, mining, roads, factories, cities, gold, and a state-centric past; “the internet” means the virtual, Anglospheric, comparatively unregulated world of AI and cryptocurrency. The recurring pairings are “state versus network,” “land versus cloud,” and gold versus digital gold.
Dave Blundin asked whether that was fundamentally centralization versus decentralization. Balaji said “roughly, yes,” but located the most interesting contests at border zones: Chinese founders choosing between integration, political suppression, or foreign startup cities; and India dividing between its physical economy and its internet-connected diaspora.
This is also Balaji’s replacement for the postwar “rules-based order.” China offers centralized sovereignty; the free internet offers decentralized contracts, speech, money, and association; America increasingly becomes contested terrain rather than a coherent third pole.
3. Blundin’s demographic objection tests the China thesis
Blundin’s pushback—worth keeping—was Japan: throughout his youth it seemed destined to overtake the US, then aging stopped the trajectory. China now has an even worse demographic profile, much lower per-capita GDP, almost no immigration, and therefore lacks the talent-renewal mechanism that sustained America.
Balaji answered first with a reversal already under way: China introduced the K visa to recruit technical talent while US nationalist politics restricted H-1Bs, researchers, students, and even tourists. He was explicit that his argument was analytical, not an endorsement of Beijing over the immigrant-capitalist America in which he built his career.
His second answer disputed Blundin’s Japan mechanism. Balaji attributes Japan’s stagnation substantially to the Plaza Accord and its subordinate relationship with Washington; China, by contrast, repeatedly accepts worse financial terms to retain control, behaving like a founder who “never gave up voting rights” even after surrendering some economics.
The Huawei handset launched during Commerce Secretary Gina Raimondo’s China visit served as his specimen: it signaled partial progress around chip restrictions and demonstrated a capacity to develop in secret. “China can do things in stealth. Japan can’t.”
4. Automation converts China’s aging problem into industrial pressure
Balaji’s central demographic counter is that “China’s demographic problems have robotic solutions.” Cheap labor can delay industrialization; a roughly 1% annual demographic shock instead forces automation in delivery, hotels, manufacturing, and other physical processes.
He urged looking past dollar-denominated GDP to physical output. His cited indicators included a Chinese shipyard producing more ships than the entire US Navy, Hegseth’s claim that Chinese hypersonics could sink all US aircraft carriers, and the Raytheon CEO’s statement that the US cannot decouple from China. He also cited a Govini study commissioned by the Pentagon that he summarized as showing that “the US military is made in China.”
On trade, Balaji argued the conflict was substantially decided during the post-2015 tariff era: only about 15% of Chinese revenue now comes from the US and 85% from elsewhere, so even a halving of the American stream leaves most revenue intact. His deliberately provocative conclusion: “The war is actually already over.”
Chip restrictions may likewise have selected for efficiency. DeepSeek and other Chinese teams pursued comparable results with fewer or less advanced chips, while China emphasized physical AI and robotics; Balaji likened challenging Chinese engineering, manufacturing, and mathematics for national pride to “challenging the Italians to a pizza-making contest.”
5. The open internet is the counterweight America should actually cultivate
Balaji is not fatalistic about Chinese predominance: “I do believe the internet can balance China.” His criticism is that MAGA displays “China envy,” competing on manufacturing and Taiwan rather than maximizing America’s distinctive advantage in open networks.
The prescription is memorable: “We have to be more America than America.” Free speech, free markets, decentralized finance, and permissionless technology are difficult for China to reproduce internally without taking apart the control architecture that gives the state its coherence.
His geographic portfolio largely inverts the 20th century. He is extremely bearish on Western Europe, bullish on Eastern Europe, moderately bullish on India but “extremely bullish on Indians,” and positive on Dubai, Riyadh, El Salvador, and perhaps Argentina—places he sees as immunized by prior experience with communism, socialism, inflation, or disorder.
6. “There are many AGIs” becomes the episode’s central wager
Balaji takes superintelligence seriously because intelligence plainly exists in many biological forms, but he thinks the observed AI market contradicts a clean hard-takeoff story. Present models excel at probabilistic, System 1-style intuition; conventional computers excel at long, explicit System 2 chains; bridging them may require another architectural “dogleg up.”
The current evidence looks polytheistic: ChatGPT, Claude, Grok, Gemini, Perplexity, Kimi, Qwen, DeepSeek, Llama 3, and others repeatedly leapfrog without one opening an orders-of-magnitude lead. Balaji praised Meta for releasing Llama 3 even while judging Llama 4 less successful.
Diamandis described a current golden era of perhaps a year or two with roughly 10 major models. Blundin argued that the natural dynamic ultimately produces one winner: true self-improvement began recently, foundation-model companies are redirecting compute from consumers into internal research, and a single dominant system could emerge within six months—two years at the outer bound.
Diamandis proposed the long bet plainly: many AGIs and ASIs versus “one digital god to rule them all.” He also argued that government involvement might preserve several players even if the natural market dynamic is winner-take-all.
7. Compute scale is no longer the only plausible winning variable
Blundin argued that transformers still contain an estimated 100–1,000x of algorithmic performance upside and asked whether an AI equivalent of OpenAI researcher Alec Radford may now be capable of inventing and implementing its own successor ideas. He cited an IQ of 148 and recent use that moved from “smart to freaking brilliant.”
Yet the competitive function keeps changing. Training-time compute once seemed decisive; inference-time compute, repeated prompting, and model-based evaluation now look increasingly important, reopening the field to AMD, new architectures, and algorithmic breakthroughs rather than only the largest NVIDIA cluster.
Blundin framed Stargate as Sam Altman’s attempt to maintain an “I’m ahead and I’m going to stay ahead” loop, with roughly $100 billion, potentially rising to $500 billion, of NVIDIA hardware and infrastructure. His honest caveat: no one inside OpenAI, Anthropic, or elsewhere knows which input wins over the next six months.
Balaji is skeptical that spending alone solves the problem: enormous investment over the preceding year and a half produced improvement, but not the radical separation incentives would predict. China’s chip-constrained efficiency and possible need for embodied, haptic feedback make a purely digital scaling race even less certain.
8. Prompting and verification remain hard bounds on autonomy
Balaji calls most AI “middle to middle,” not end to end. A prompt is like a high-dimensional vector pointing a fantastically fast spacecraft; a human still chooses the vector, and another human or external system must verify where it lands.
Self-driving is his counterexample: it finally performs an entire trip, but only after roughly 20 years since the DARPA Grand Challenge, billions of dollars, millions of miles, and extensive physical-world transition work. That history argues against instant end-to-end replacement across every occupation.
His deeper objection to omniscience is mathematical. “Crypto is what AI can’t do. Chaos is what AI can’t do. Turbulence is what AI can’t do”: cryptographic preimages, chaotic systems, and physically turbulent randomness impose limits that, in his view, no amount of simulated contemplation can simply forecast through.
Markets, politics, and media add time-varying adversaries. A trade or post that worked yesterday changes the environment once deployed, while competing AIs immediately respond; Balaji sees that reflexivity as both a constraint on prediction and a force decentralizing capability.
9. Blundin and Balaji split over whether evaluation loops close the gap
Blundin’s strongest case is massive parallelism: AI can attempt 10,000 solutions concurrently, use other models to score them, and preserve the best. With reliable evaluators, that loop can accelerate code, games, movies, physics, mathematics, and potentially AI research itself.
Balaji’s reply invokes Goodhart’s law: once a benchmark becomes the target, systems learn the benchmark rather than the high-dimensional surface users actually care about. Grok’s benchmark strength, he said, does not make it obviously superior in daily use to Claude; an isolated metric is merely “a hair on a very large head.”
Verification is also costly. Midjourney may enter the right neighborhood quickly but require hundreds or thousands of samples to find the desired image; probabilistic chains compound errors like a chemical process yielding 95% at each of 20 steps.
Diamandis noted that internal researchers can freeze random seeds, set temperature to zero, and reproduce tests exactly. Balaji accepted that distinction but maintained that outside users face stochastic output, making unsupervised public posting or financial action dangerous: one bad like, trade, or transfer remains the owner’s responsibility.
10. AI amplifies the operator before it replaces the operator
Balaji’s signature framing is “amplified intelligence, not artificial intelligence.” A mediocre prompt produces polished “AI slop”—his joke was “lorem AI ipsum”—whereas a domain expert can explore valuable corners and distinguish a discovery from confident gibberish.
Terry Tao was his model user: someone with enough mathematical knowledge to parse a stream of unfamiliar symbols and verify whether it contains a real idea. Without that expertise, a user can be persuaded that the model discovered new quantum mechanics when it merely generated plausible language.
Hence “everybody’s a CEO, everybody’s a manager.” AI resembles a talented junior employee whose work must be checked; it needs to satisfy the manager, but the founder still faces “the unforgiving market,” which can mark an unverified product directly to zero.
Models also compete for an already-created budget category: “AI doesn’t take your job. It takes the job of the previous AI.” Firms will swap image, code, text, and video systems, while permanent AI teams optimize workflows and new verification, proctoring, and proof-of-human services police the resulting spam and scams.
11. Generalists regain wingspan, but expertise still finishes the work
AI can raise a novice to “a six or a seven” in design, art, coding, or prototyping, which is transformative for founders and people rich in time but short on money. Production polish still usually requires an expert capable of debugging the relevant symbols.
The resulting economy reverses “the extreme specialization of the 20th century.” Balaji invoked the line “specialization is for insects”: individuals and small teams can write, design, prototype, research, and operate across disciplines, while smaller communities become more technologically self-sufficient.
Some template-driven work remains directly exposed. AI may diagnose more cheaply, quickly, personally, and sometimes with fewer errors; patients can obtain a “20th opinion,” and clients can draft contracts before a professional signs off. The surviving MD or JD function becomes final certification—protected by pressure from the AMA, ABA, and other “white-collar unions.”
12. Biomedical AI moves fastest where reality supplies an evaluator
The speakers agreed that fields progress at different rates according to two variables: whether outputs can be objectively evaluated and whether regulation permits deployment. Coding, mathematics, and parts of physics face few legal barriers; medicine and law combine excellent evaluation opportunities with powerful gatekeepers.
Balaji’s clean specimen was protein folding: primary sequence goes in, a 3D structure can be checked against the PDB and X-ray crystallography, producing a known objective function. Balaji also cited high-resolution medical scans, where machines can inspect terabytes no radiologist has time to examine and still return a visibly checkable finding.
Balaji is especially bullish on biomedical text mining—reconciling papers, references, contradictions, and relationships around targets such as p53—and on mixed text-image diagnosis. But he distinguishes synthesis among existing nodes from discovering points outside the known envelope: interpolation can be genuinely useful without amounting to new physics or biology.
Diamandis asked about Demis Hassabis’s forecast of curing all disease by 2035 and Dario Amodei’s suggestion that lifespan might double within a decade. Balaji did not reject the technology; he rejected analyses that omit the institutions controlling what reaches a patient.
13. Regulation, not model intelligence, may be longevity’s binding constraint
Balaji’s hierarchy is political: “AI is below the state and crypto is above the state.” Bitcoin and smart contracts can replace or politically challenge parts of the Fed and SEC; biomedical AI still depends on FDA approval, CPT reimbursement, medical licensing, prescriptions, and a functional state that many AI forecasts simply assume.
After more than 20 years around genomics and biotech, he identifies the regulatory state as the primary blocker. His historical contrast was insulin: Banting and Best moved from dogs to themselves and willing patients, then scaled with Eli Lilly between roughly 1921 and 1923—“pharma moved at the speed of software” before modern phased trials.
Balaji illustrated the asymmetry in tort law: one treatment error can produce a $100 million settlement, while a thousand deaths from a delayed cure create no equivalent liability. He generalized it: inside large institutions, “risk budget is always scarcer than budget.”
Startup jurisdictions could change the equation through voluntary risk. Balaji cited Próspera and Minicircle’s human gene-therapy work; Diamandis pushed back that he was still waiting for published science rather than testimonials. Both preserved geographic arbitrage as the mechanism: successful treatments might begin in El Salvador, special economic zones, or longevity-tourism hubs before larger regulators move.
14. Bitcoin and longevity reject managed decline on the same grounds
Balaji’s deepest connection between the two is moral, not transactional. Traditional finance says hyperinflation is bad but roughly 2% annual currency debasement is healthy; traditional medicine says rapid death is bad but reversing aging is strange. One asks people to lose “a little bit of wealth,” the other “a little bit of health,” every year.
His first-principles challenge is “Why is there a hole in our bucket?” Bowhead whales reportedly live around 200 years and Greenland sharks around 500, making repair of human biological software and hardware, in his view, “not a matter of if, it’s only a matter of when.”
Progress therefore needs “medical heroes” and risk-tolerant jurisdictions. “No plane crashes, no planes. No train crashes, no trains”: Balaji’s darkly comic example was the old CRC Handbook listing compounds’ smells and tastes until researchers learned, sometimes personally, which exposures were dangerous.
15. Bitcoin is settlement money; its proxies can power the machine economy
Balaji compares Bitcoin to a power station: “really high-voltage money” transferred relatively infrequently, with lower-voltage representations serving everyday use. He places base-layer transaction volume near Fedwire and calls it “the Fedwire of the free world.”
Scaling need not be purely peer to peer. Coinbase and Binance can handle rapid internal transfers and settle net balances between hubs in BTC; many Lightning implementations, Balaji said, are in practice similar peering agreements between applications. Users retain the option to withdraw, even if this resembles a hub-and-spoke network more than a perfectly decentralized retail system.
Base BTC therefore may not become the direct currency for every AI action. Diamandis argued that pure BTC transaction volume is insufficient for all AI payments; wrapped BTC, off-chain claims, or BTC on another chain—he specifically mentioned wrapped BTC on Solana—can support high-throughput payments while the underlying asset remains the scarce reserve.
Asked about his allocation, Balaji declined to provide a percentage but answered: “Only hold in USD what you can afford to lose.” Peter separately said that, if forced into fiat, he would prefer Singapore dollars or UAE dirhams; he rejected CHF’s former safe-haven status and was negative on Western currencies, JPY, and probably KRW.
16. Currency competition moves from geography to software features
Balaji expects states to settle more trade with physical gold while Bitcoin becomes “the reserve currency of the network.” Gold’s security costs suit governments; digital gold suits online actors, and both appreciate against Western fiat in his model.
His newspaper analogy supplies the causal chain: local papers survived because trucks created geographic monopolies; once news went online, national and then internet-native competitors erased them. Likewise, fiat currencies first move on-chain, lose their captive geographic users, and compete globally on ideology and features.
Most local currencies then disappear, perhaps leaving USD, RMB, and a few others before all devalue against BTC. The premise is simple: once a wallet can hold the best available currency, “why do you have to hold the currency of your geography again?”
17. Crypto agents, network states, and a political backlash complete the thesis
Crypto is the obvious machine rail because software can create unlimited wallets, hold private keys, and fund a robot in fractions of a second. Transactions that are “very large, very small, very fast, very automatic, very international, very transparent or very complex” expose the limits of legacy banking.
Balaji says crypto exchanges are becoming the new banks through cross-border peering. Blundin’s pushback was that banks do more than custody M1: consumer lending, credit evaluation, and other services remain. Balaji narrowed his disruption claim to retail banking and power users, for whom much of the migration has already happened.
The repricing mechanism is the “billionaire flippening.” Balaji and Polychain’s Olaf Carlson-Wee estimated that somewhere between $100,000 and $1 million per BTC, half of global billionaires become crypto billionaires. Balaji said another 10x would dilute fiat billionaires, banks, and states; Diamandis illustrated the portfolio effect with a previously balanced $1 million crypto/$1 million non-crypto portfolio becoming roughly 90–95% crypto after a 10x move.
Bitcoin can still fail individually. Quantum risk may require quantum-safe encryption and a difficult protocol change, possibly a hard fork, to move funds, but Balaji argues crypto as a class now has “herd immunity”: proof-of-work and proof-of-stake systems fail differently, chains can anchor one another, and attacked assets can escape through snapshots and “lifeboats.”
Network states are the institutional extension. Balaji moved from the US in 2020 and describes the internet as “America 2.0”: a place where globally overlooked “dark talent” receives equal contract and monetary rights. His Network School is funded from the “Bank of Balaji,” recruits wherever English is spoken, and aims to develop founders, workers, currencies, and durable physical communities.
The ambition moves from internet company to internet currency to internet community. Balaji noted that roughly 5.6–6% of Thiel Fellows became unicorn founders, but communities matter beyond rare unicorns: valuing Singapore by its land or residents produces an estimate around $10 trillion, South Korea around $100 trillion, and China-scale startup societies near a quadrillion dollars.
His practical advice starts with “go direct”: founders should build their own audiences, keep content creation as close to the company as code, use AI without outsourcing their authentic voice, and avoid legacy-media dependence. In the US he favors Texas or Florida, cold-storage self-custody, and trusted offline tech communities, including microschools and other collective institutions.
Individuals with mathematical aptitude should take a focused month offline to learn computer science, statistics, gradient descent, Karpathy’s material, fast.ai, and perhaps crypto. Balaji pairs extreme connectivity with Faraday-cage-style periods of pencil, paper, coffee, and quiet: “The better you can focus offline, the stronger you’ll be online.”
His capital discipline is long-term and anti-emotional: understand crypto without treating the comment as personalized advice, research assets after they leave the headlines, use 30-to-150-day calendar reminders, write down the thesis, and avoid day trading. “Don’t gamble”; build, buy and hold for the long run where appropriate, and respect other builders rather than scoring cheap attacks online.
The political warning beneath the checklist is that fewer than 4% of US marriages reportedly cross the Democrat-Republican divide, and Balaji expects the 2030s to bring “red and blue against tech.” AI job fears, data-center electricity, crypto, biotech and IVF, offshore production, and visible founder wealth create a common target; his hopeful close was that the internet can still deliver abundance after “some of the white water that’s coming up.”