Balaji Srinivasan: How AI Will Change Politics, War, and Money
Technologist Balaji Srinivasan joins a16z's Erik Torenberg and Martin Casado to discuss the limitations and societal impact of AI, framing the conversation around the concept of "Polytheistic AGI"—multiple, culturally-specific AIs—versus a singular, god-like intelligence. They explore the practical system-level constraints on AI, its surprising evolution, the critical role of cryptography in grounding AI in reality, and the future of work and security in an AI-driven world.
The discourse around AI is often framed by an implicit “monotheistic” vision of a single, all-powerful AGI, a concept that evokes both awe and fear, much like an Abrahamic god. An alternative and more practical framework is that of “Polytheistic AGI,” which posits the emergence of many superhuman intelligences, each shaped by different cultural values and contexts. This suggests a future with not one, but multiple AIs—an American AI, a Chinese AI, and hopefully, decentralized open-source AIs.
In this polytheistic world, each society or “network state” might have a core technological stack: a social network for community, a cryptocurrency for deterministic law, and a customized AI to serve as a probabilistic oracle or guide.
The Platonic Ideal vs. System Realities
Much of the fear surrounding AI stems from conflating Platonic thought experiments, like those from Nick Bostrom’s Superintelligence, with the real-world computer systems we have today. While these ideals are useful for framing conversations, it’s crucial to ground the discussion in the known limitations of computer science.
AI models are not gods; they are software running on hardware and are bound by fundamental constraints.
- Chaos and Unpredictability: AI cannot predict everything. Systems governed by chaos, turbulence, or cryptography are inherently resistant to prediction. A small change in initial conditions creates a vastly different outcome, placing a hard mathematical bound on an AI’s forecasting ability.
- Closing the Control Loop: The idea of an AI “jumping out of the box” and recursively self-improving hasn’t materialized because of the difficulty in “closing the loop.” An AI doesn’t inherently know what it doesn’t know. If it generates an output (a new “direction”) and feeds it back into itself, there’s no guarantee that the new direction is “in-distribution” or coherent, leading to error accumulation. Self-reflection is a key missing component for true autonomy.
Prompting, Verification, and Embodiment
The limitations of AI autonomy are evident in the human-centric processes still required to operate them. AI currently operates “middle-to-middle,” not “end-to-end.”
- Prompting as a High-Dimensional Program: A prompt is more than a simple query; it’s a high-dimensional vector pointing the AI “spaceship” in a specific direction. The art of prompting is akin to writing a tiny program for an undocumented but highly error-tolerant API. A richer vocabulary and deeper subject matter expertise allow a user to generate superior results, amplifying the intelligence of the expert.
- The Primacy of Verification: While prompting gets attention, verification is the more critical and expensive step. AI can generate vast amounts of content, but this necessitates a massive human effort in “proctoring and verification” to ensure its accuracy and safety, creating a new category of jobs.
- Lack of Embodiment: Current AIs are not embodied. They cannot script robots to build their own data centers or mines. This physical constraint is a primary reason they lack true goal-setting, reproduction, and independent action.
AI’s Domains: Where It Excels and Fails
AI’s capabilities are not uniform across all domains. Its performance is surprisingly counterintuitive.
- Visual over Verbal: AI is more effective at visual and front-end tasks (generating images, UIs) than back-end or logical tasks (writing complex code, legal contracts). The reason is verification cost. A human can instantly and intuitively (System 1 thinking) assess the quality of an image. Verifying code, however, requires slow, deliberate, line-by-line analysis (System 2 thinking).
- Stateless over Stateful: This distinction extends to stateless versus stateful systems. An image is stateless—all its information is present at once. A piece of code is stateful; its behavior evolves at runtime and is often computationally irreducible, meaning it must be run to be fully understood.
- Poor Performance in Adversarial Systems: AI struggles with markets and politics. Unlike games with fixed rules like chess, these domains are time-varying, rule-varying, and adversarial. Any successful strategy is quickly identified and countered by other actors (many of whom are also using AI), making static models ineffective. The human “sensor” who understands the shifting dynamics remains essential.
AI, Crypto, and Grounding in Reality
A key challenge for AI is the “grounding problem”—connecting its outputs to verifiable reality.
AI makes everything fake and crypto makes it real again.
AI is a probabilistic technology, excellent at generating plausible fictions. Cryptography, conversely, is a deterministic technology that provides hard, verifiable truths. It represents what AI cannot fake, such as a private key or an on-chain transaction.
By citing on-chain data—from block explorers, crypto instruments, or cryptographically signed posts—an AI can ground its assertions in a verifiable, tamper-proof reality. While this doesn’t solve the initial data ingest problem (i.e., proving a photo wasn’t doctored before being hashed on-chain), it creates a chain of provenance and makes coordinated deception far more difficult.
The Future of Work and Society
AI’s impact will be transformative, but not in the way many expect.
- Amplified Intelligence: AI primarily acts as a force multiplier, making experts more effective. Senior developers, for instance, see greater productivity gains than junior ones because they are better equipped to prompt, interpret, and verify the AI’s output.
- Job Competition: AI doesn’t take a human’s job so much as it takes the job of the previous AI. Companies now have a “slot on the roster” for an AI image editor or code assistant, and new models from Google, Anthropic, and others compete for that slot.
- Specialists and Generalists: AI empowers generalists to achieve a baseline level of competency in many fields. However, specialists are still required for true polish and to unlock the highest-level capabilities of the models through expert prompting.
- The Anti-AI Backlash: A significant backlash is inevitable, driven by both economic and political forces. Media unions are already trying to contractually block the use of AI to protect jobs, and politicians are beginning to use AI as a powerful tool to mobilize voters by tapping into deep-seated fears about technology.
- Security and Geopolitics: The most immediate threat is not a “super-persuader” AI but “killer AI” in the form of autonomous drones, which are already being pursued by every major country. This makes the concept of “digital borders” a physical reality, potentially re-asserting the power of the nation-state to control its territory from digital and physical incursions.