Imagine a future where every distinct culture, ideology, and community is guided by its own artificial intelligence—an oracle reflecting its values, integrated with its own social network and cryptocurrency. This vision of “polytheistic AGI,” where we see a multitude of superhuman intelligences rather than a single, monolithic god-like AI, frames a wide-ranging conversation between technologists Balaji Srinivasan and Martin Casado. They dissect how our language and metaphors for AI—casting it as a deity, a tool, or a swarm—shape both our ambitions and our anxieties. Moving beyond the hype, they explore the tangible limitations of current systems, the cultural imprinting inevitable in AI development, and the practical reality that these technologies are, at their core, software bound by computational laws.

The discussion rigorously challenges the “platonic ideal” of AI that has dominated public discourse, inherited from philosophical thought experiments like Nick Bostrom’s. Casado, a systems expert, grounds the conversation in the concrete constraints of computer science: AIs cannot simulate chaotic or turbulent systems with perfect long-term predictability, nor can they easily “close the control loop” to operate fully autonomously. A key insight is that today’s AI operates “middle to middle”; it requires a human in the loop to provide the high-dimensional direction (the prompt) and, crucially, to verify its outputs. This makes the technology more of an amplifier than an autonomous agent, disproportionately benefiting those with existing expertise who can ask better questions and better judge the answers.

The conversation also delves into the emerging geopolitical and social landscape shaped by these tools. The idea of polytheistic AGI suggests a world where American, Chinese, and decentralized open-source AIs will evolve with different embedded values and restrictions, acting as the core of modern “network states.” This plurality inherently mitigates some existential risk scenarios. Furthermore, they posit that the true “killer AI” is not a persuasive chatbot but embodied, drone-based warfare, which is already being developed and fundamentally alters concepts of borders and power. The backlash against AI, they predict, will be profound, not just from displaced workers but as a potent political tool that taps into deep-seated human fears about technology and agency.

Surprising Insights

• AI excels at “hard” human tasks (like writing sonnets) but struggles with “easy” ones (like systems programming or robotics). This inverts our intuition because humans find creative tasks difficult, but we’ve had millions of years of evolution perfecting physical navigation and interaction with chaotic, real-world environments.
• The most significant limitation for autonomous AI is “closing the control loop.” An AI generating its own next prompt is akin to a spaceship needing to plot its own course in a near-infinite dimensional space; it doesn’t know what it doesn’t know, leading it quickly out of its trained distribution and into nonsense.
• Cryptography (via blockchain) is framed as the deterministic antidote to probabilistic AI fakery. While AI can generate limitless synthetic content, it cannot forge a valid on-chain cryptocurrency transaction or a cryptographically-signed piece of data, creating a hard anchor for truth in a world of easy simulation.
• AI doesn’t primarily replace human jobs; it replaces previous AI jobs. Within an organization, each new, more powerful model (like Claude or Grok) competes to take the “seat” held by the prior model (like ChatGPT), while humans shift to higher-level prompting, proctoring, and verification roles.
• Global wage convergence may be accelerated by AI. A lawyer in the West and a similarly intelligent person in a developing country, both augmented by the same AI tools, could see their productive output—and thus their wages—converge, creating massive disruption and consumer surplus.

Practical Takeaways

• Invest in “verification” skills. As AI gets better at generating content, code, and data, the premium shifts to the human ability to audit, check, and validate that output. Developing a keen eye for detail and error will be a critical career skill.
• Treat prompting as a high-level programming language. The quality of an AI’s output is directly tied to the precision and richness of your prompt. Building a broad vocabulary and specific domain knowledge allows you to craft these “tiny programs” more effectively.
• Use multiple AI models like consulting a pantheon of experts. Don’t rely on a single model. Present important questions to different AIs (ChatGPT, Claude, etc.), let them argue with each other, and synthesize the answers yourself to mitigate individual biases and errors.
• Focus on skills that involve navigating uncertainty and adversarial systems. AI is currently weak at time-varying, rule-varying, and adversarial domains like market speculation or politics. Human intuition, strategy, and real-time sensing will remain valuable in these areas.
• For complex tasks, use AI for the visual and human verification for the verbal/logical. It’s easier for humans to instantly verify an image, UI, or graph. Use AI for that. For code, legal text, or math, assume you must slow down and verify each step meticulously, as errors are harder to spot instantly.