Balaji Srinivasan: How AI Will Change Politics, War, and Money

Summary

• Srinivasan’s core macro call is “polytheistic AGI”: not one unitary intelligence taking off to infinity, but American, Chinese, and decentralized open-source models multiplying into culturally specific systems. Each internet-first society could combine AI as its probabilistic “oracle,” crypto as deterministic law, and a social network as connective tissue—the “reactor core” of a network state—making model plurality and customization more consequential than a single AGI winner.
• The AI-apocalypse framing mistakes a Platonic ideal for software bounded by computation, chaos, turbulence, and cryptography. Casado stresses that today’s models are real computer systems, while Srinivasan notes that the fast-takeoff scenario did not occur and current AI lacks goal-setting, embodiment, reproduction, and independent action—although he treats self-replication as a constraint “today,” not necessarily forever.
• The emerging AI economy is “middle to middle”: humans supply direction through prompts and then absorb the expensive work of verification. A prompt is a high-dimensional heading for a fast spaceship, while autonomous feedback is fragile because the model “doesn’t know what it knows” and is “optimized to fake it.” Srinivasan expects business spending and employment to move toward prompting, proctoring, and verification; Torenberg likens the broader verification burden to KYC.
• Near-term value accrues fastest in visual, front-end, and stateless work whose quality can be inspected almost instantly. Images, video, and interfaces expose their gestalt cheaply; backend code, legal language, and mathematics require slow System 2 review, while stateful software may be computationally irreducible. “AI makes everything fake,” Srinivasan argues, so verification tooling becomes part of the product.
• AI appears more likely to amplify expertise than erase its advantage. Casado says early coding data show senior developers receiving larger relative productivity gains because they ask better questions, recognize trade-offs, and reject bad output; Srinivasan calls this “amplified intelligence, not agentic intelligence.” AI can make “everyone a CEO,” but the sharp formulation is that it often “takes the job of the previous AI.”
• Markets and politics remain hostile domains because they are time-varying, rule-varying, and adversarial. A strategy degrades once competitors adopt it, and “the other guys are also using an AI on you,” leaving the CEO, influencer, or creator as the live sensor that interprets changing conditions. Srinivasan keeps StarCraft as a genuine counterexample that complicates, but does not erase, the boundary.
• Crypto can authenticate digital history, but Casado’s pushback is that it cannot by itself prove the physical-world input was true. Srinivasan’s answer runs from FTX transfers verified on block explorers to Farcaster posts, crypto IDs, and instruments that hash camera or sequencing-machine output onchain at capture time. That creates stronger provenance and coordinated attestations, “not impossible” forgery resistance.
• The most concrete “killer AI” is already drones, with downstream consequences for borders, surveillance, and political backlash. Srinivasan expects digital borders to harden as remote systems can control machines inside a jurisdiction, while AI makes previously unsearchable surveillance archives queryable. Labor pressure compounds the politics: his illustrative convergence moves a Western professional from $200K toward $20K while lifting a $2,000 overseas worker tenfold.

Deep dive

1. AGI is becoming plural before it becomes godlike

• Torenberg’s opening invites an intellectual history: Srinivasan earned his Stanford PhD around 2005–06, taught machine learning and computational statistics through genomics, founded a DNA-sequencing company, and spent roughly a decade doing ML full-time before crypto absorbed him during the early-2010s deep-learning turn.

• Srinivasan’s honest revision: despite following ImageNet, diffusion models, style transfer, GPT-2, and related language models, he expected language generation to remain “Markov chain-like.” GPT-3 and DALL-E in early 2022 were a hint, but ChatGPT’s coherence was a “huge jump”; even predict-next-token systems and double descent traveled much farther than his classical-ML intuition expected.

• The monotheistic AGI story, in his telling, implicitly seeks a unitary intelligence that reaches infinity—part Abrahamic God, part rapture, part paperclip-producing avenger. His alternative is “war of the gods”: many superhuman systems trained within different cultures, carrying different values, restrictions, and notions of acceptable speech or imagery.

• In 2022, Srinivasan expected at least American AI and Chinese AI and, “if we’re lucky,” decentralized open-source AI. DeepSeek and the near-weekly arrival of strong open models now make plurality clearer: each culture could eventually pair its AI “oracle” with cryptocurrency and a social network, forming the “reactor core” of an internet-first society.

2. Software limits puncture the god metaphor

• Casado’s essential distinction is between a thought experiment and an implemented system. Nick Bostrom’s superintelligence could recursively self-improve because the premise granted it that power; importing those assumptions into LLMs arriving four years later obscured known bounds on computer simulation, time, precision, and compute.

• Casado defends Platonic ideals with examples: the Turing test moved from thought experiment toward applied systems via the Lovelace test and CAPTCHA v0; the Chinese room is another implementation-agnostic thought experiment; and social networks made six degrees of separation real in practice. Their shared resolution is methodological—keep the thought experiment, but state clearly when it stops describing actual software.

• Prediction itself has hard boundaries. Turbulence and chaotic systems fracture under finite-precision forecasts, while an MD5-style hash turns a one-character change into a radically different result; Srinivasan’s physical thought experiment is to introduce turbulence before throwing a pitch, making the decision path impossible to forecast indefinitely.

• Casado explains the locomotion paradox through evolutionary “equilibrium”: spatial navigation competes with a roughly 4-million-year-old mammalian system, while the prefrontal cortex handling language learning and creativity was described as about 250,000 years old. Srinivasan remains struck by how much world structure language contains; Casado’s qualification is that humans may have first constructed that world model, then cached it in text.

3. Autonomous AI still fails at direction and feedback

• Srinivasan treats self-replication as a present limitation, not a permanent theorem. Today’s AI lacks embodiment, independent goal-setting, reproduction, and the ability to build its own mines or data centers; his consciousness analogy is an animal modeling whether its body will fit beneath a branch while escaping a predator.

• A prompt is “a very high-dimensional direction vector.” Even a near-light-speed spaceship must be aimed with two coordinates and successive waypoints; an AI receives far richer symbolic coordinates, so speed or intelligence does not eliminate the need to select a destination. “Prompts are tiny programs,” but they target a hidden, undocumented, highly error-tolerant API.

• Casado adds the harder problem: closing the control loop. A model generating its next instruction may wander out of distribution because it “doesn’t know what it knows and it doesn’t know what it doesn’t know”; feeding that output back can compound error. The model can list thin areas—real-time events, niche fields, paywalled work, local information, human intent, proprietary systems—but cannot guarantee safe self-direction.

4. Verification captures the spend that generation removes

• Srinivasan’s commercial formulation is crisp: “AI doesn’t do it end to end. It does it middle to middle.” People still prompt at the start and verify at the finish, and a conversation with Andrej Karpathy led him to expect “massive numbers of jobs” in proctoring and verification precisely because models generate plausible fakes so cheaply.

• Visual output has a structural advantage: images, video, and interfaces from tools such as Vercel’s v0 or Replit can be judged almost immediately with human perception and available GPUs. Backend code, legal language, and equations require line-by-line System 2 reasoning; the cost AI removes from production may simply reappear in inspection.

• Casado sharpens the split into stateless versus stateful systems. An image exposes essentially all its state, whereas a short backend routine can conceal a complex runtime state machine; some behavior must actually be executed because it is computationally irreducible. Formal verification is viable for small, high-value smart contracts, not general software.

• Their prospective AI UX makes internal failure perceptible: turn audio into spectrograms, sonify model behavior so anomalies resemble a car’s unfamiliar rumble, or color text by confidence. The systems implication is that tool use lets a fuzzy model call deterministic software; Srinivasan treats that as evidence that traditional systems still do the precise work, while Casado leaves open whether a pragmatic hybrid could nevertheless get there.

5. Crypto authenticates history after reality enters the system

• Srinivasan’s paired slogan is “AI makes everything fake and crypto makes it real again.” Probabilistic generation cannot counterfeit a Bitcoin private key, valid digital signature, or onchain NFT; cryptographic equations provide the hard boundary that generated text, images, and identities lack.

• His FTX example grounds the claim: a Perplexity summary of the 2022 hack could cite a block explorer, letting the reader verify transfers, signatures, hashes, and timestamps. As block space expands, Farcaster-style posts and crypto IDs could extend those citations from financial data to social assertions and tamper-evident metadata.

• Casado’s pushback—worth keeping—is the data-ingest problem. An on-chain signature record can prove that a key made an assertion at a timestamp, but not whether a claimed photograph, location, or physical event was genuine. Crypto offers end-to-end guarantees “as soon as you get it into the system”; it does not automatically ground the system in reality.

• Srinivasan’s partial answer is the “crypto instrument”: hash camera data or DNA-sequencing TIFF files as they are captured, timestamp them onchain, and combine them with preregistered trials and proof-of-human attestations. Coordinated fakery becomes harder, “not impossible,” and Casado agrees rising incentives should improve the truthfulness of inputs over time.

6. Markets and politics resist frozen models

• Static mappings—cat labels, chess rules, checkers, or Go—fit train-and-test methods. Markets and politics do not: they are “time varying, especially rule varying and adversarial,” the same trade eventually becomes a loss, and competitors deploy AI too. Torenberg therefore casts the CEO, influencer, or creator as the sensor reading live incentives and human nature.

• Casado connects that boundary to nonlinear, chaotic systems where extrapolation is intrinsically difficult. Srinivasan accepts the connection but preserves an awkward counterexample: AI performs well at StarCraft, which is adversarial and more time-varying than chess, although not genuinely rule-varying. The distinction is a working hypothesis, not a claimed universal law.

7. AI compounds expertise while models compete with models

• “AI means amplified intelligence, not agentic intelligence” because stronger writers, engineers, and domain experts can formulate better instructions and verify the result. Casado says coding productivity data already show senior developers gaining more even on a relative basis: they understand trade-offs, speak formal languages efficiently, and know what to discard.

• Srinivasan’s managerial extension is that “AI means everyone’s a CEO”: giving clear written instructions and reviewing execution dramatically lowers the cost of managing experimental projects. Yet each company also develops fixed roster slots for image, text, and coding models, so a new Claude, Grok, or ChatGPT often “takes the job of the previous AI.”

• The expert and casual-user markets can coexist. A programmer may request a serviceable 3D asset without learning design, while a specialist extracts better output through domain vocabulary; similarly, Lovable serves casual website creation, whereas Cursor wraps AI inside the professional developer’s IDE. Casual competence does not erase the specialist’s advantage in polish.

• Casado challenges polytheism with rapid distillation: leading models teach rivals, even converging on the same supposedly random number. Srinivasan imagines a shared “spinal column” with cultural differentiation above it; Srinivasan also says pre-training broadly improved models, whereas current debate and data suggest domain-specific RL does not generalize in the same way and may “rob Peter to pay Paul,” creating lasting demand for specialized models around fundamental trade-offs.

8. Drones turn AI from persuasion into territorial power

• Srinivasan’s blunt correction to safety discourse is that “killer AI is already here and it’s called drones,” a point Casado strongly endorses. Image generators and speculative “super persuaders” receive attention while every country pursues weaponized automation; the rhetoric shifted from safety to beating China, but both frames can justify control.

• Digital borders could therefore become hard borders. China already describes the Great Firewall as packet-level sovereignty, and remote drone or humanoid control makes “cloud space” physical; Ukraine’s multi-kilometer, cable-controlled one-way drones illustrate how weapons can bypass jamming. Casado’s counterpoint is decisive: a fully autonomous drone given a target image may need no incoming packets at all.

• Srinivasan contrasts defensible territory with an “encrypted state.” A country outlined on a map is targetable, while Bitcoin has no complete, current map of every holder or miner; dispersion creates a degree of “security through obscurity.” Future sovereignty may split between jurisdictions enforcing hard digital borders and networks whose assets cannot be geographically enumerated.

9. AI backlash becomes a labor and power struggle

• Srinivasan’s China anecdote supplies the surveillance baseline: while recounting his movements on a call, repeated disconnections culminated when he said “Tiananmen Square.” Srinivasan argues AI changes scale, turning years of video and communications into searchable history; the old “mountains are high and the emperor is far away” gives way to Torenberg’s “long arm” and Srinivasan’s claim that it is “infinite.” Cryptography, property resistant to seizure, and exit become potential checks.

• Torenberg describes the backlash as already institutional. Media unions are writing contracts that prevent owners or editors from using AI, which Srinivasan thinks makes incumbents brittle against AI-enabled competitors; artist forums divide people into supporters and opponents. Casado notes AI artists may invest comparable effort through different tools, while Torenberg compares the social reaction to master craftsmen confronting 1800s mass production.

• Global labor arbitrage supplies the material grievance. Srinivasan’s example pairs a $200K Western lawyer or doctor with a $2,000 worker in India or the Philippines; AI plus human capability might converge near $20K, giving the overseas worker a 10x gain while cutting the Western wage to one-tenth and expanding consumer surplus.

• Srinivasan says, perhaps cynically, that politics involves sophisticated people serving clientele classes and choosing sound bites that move them. Torenberg adds that patrons can hold nuance but deliberately dumb it down, making AI a potent political tool on both the right and left.

• Casado closes by grouping the disruption: AI in media, crypto over money, robots over manufacturing, and drones over military power. He also sees crypto angles in money, AI’s constraints, and an on-chain drone control plane, while saying godlike AI taps a core human insecurity present in myth and legend for 3,000 years. Together, they frame the anti-AI fight as part of the wider anti-tech backlash, not merely a dispute over chatbots.