AI Doesn't Need to Understand Politics to Upend the World: Technological Generational Gaps Are the Real Lever of Change

A Thought Experiment from the 18th Century

AI safety researcher Ryan Greenblatt proposed a strikingly powerful analogy: if you wanted to change the world in the 18th century, one approach would be to learn how to navigate and master the politics of Westminster. But there's an entirely different approach — you could simply start building steamships, telegraphs, and Maxim guns.

In other words, if you traveled back to the 18th century with modern engineering capabilities, you wouldn't need to understand the political dynamics of that era at all. You could simply say: "I have no idea what you're debating in Parliament, but I have a fleet of steamships and a stockpile of Maxim guns." With such an overwhelming technological advantage, you would naturally become the most disruptive force of that age.

The choice of these three technologies was not arbitrary. The steamship (commercially realized with Fulton's Clermont in 1807) fundamentally transformed global trade patterns and military power projection. The telegraph (invented by Morse in 1837) compressed information transmission from weeks to seconds, reshaping how empires were governed. The Maxim gun (invented in 1884) created overwhelming firepower advantages, enabling a handful of industrialized nations to control vast colonies with minimal troops. What these three technologies share in common is that they didn't require social consensus or political permission to produce crushing effects — they altered the power landscape purely through physical capability advantages.

The core insight of this analogy is: the ability to change the world doesn't necessarily come from social maneuvering — it can come from sheer technological dominance.

AI's "Verifiable Capabilities" Are Advancing Rapidly

Greenblatt maps this logic onto current AI development trends. He points out that AI's capabilities on "verifiable stuff" are improving rapidly — chip design being a prime example. These domains share a common trait: the quality of results can be objectively measured and verified, making them ideally suited for AI to continuously improve through large-scale training and iteration.

The concept of "verifiable tasks" has deep theoretical roots in machine learning. The core premise of Reinforcement Learning is the need for a quantifiable reward signal to guide model optimization. In chip design, metrics like power consumption, area, and timing can be precisely simulated and scored using EDA (Electronic Design Automation) tools. In code writing, whether a program passes test cases is similarly a binary verifiable outcome. In 2022, Google DeepMind demonstrated AI's ability to design chip layouts (floorplanning) with AlphaChip, producing layouts that matched or even surpassed senior engineers on key metrics. The defining characteristic of these tasks is the existence of a clear objective function, allowing AI to continuously self-improve through massive trial-and-error feedback loops at a pace far exceeding domains that rely on ambiguous human feedback.

This stands in stark contrast to AI's performance on tasks like "playing politics." Political maneuvering involves ambiguous social consensus, hard-to-quantify interpersonal relationships, and complex value judgments — domains where AI progress has been much slower. But Greenblatt's point is precisely this: AI doesn't need to be good at politics to bring about disruptive change.

As long as AI reaches a sufficiently high level in hardcore engineering domains like hardware R&D, robotics, chip design, and even building fabs, it becomes the equivalent of that time traveler wielding steamships and Maxim guns.

Technological Generational Gaps Are the Ultimate Lever of Change

The key logic here is: truly world-changing transformation typically stems from absolute generational gaps in capability, not social finesse.

Greenblatt argues that if AI becomes sufficiently powerful in R&D — especially hardware R&D and robotics — it could fundamentally reshape the world, even if it has zero understanding of political manipulation. This kind of transformation isn't achieved through persuading humans, winning debates, or controlling narratives. It's achieved by directly building physical and economic infrastructure that far surpasses current technological levels.

The central role of fabs in this logic deserves special attention. Modern advanced-node fabs (such as TSMC's 3nm/2nm facilities) cost over $20 billion, take 3–5 years to build, and require the coordination of thousands of precision tools within extremely clean environments. Only a handful of locations worldwide can manufacture the most advanced chips, making semiconductor manufacturing capability a core bargaining chip in geopolitical competition. If AI could optimize or even autonomously design fab processes — from lithography parameter tuning to yield improvement — its impact would far exceed anything achievable through political negotiation. Even more concerning, this essentially represents control over compute production capacity, and compute is the material foundation for AI's own further development, creating a potential self-accelerating loop.

Just as 18th-century parliamentary debates became irrelevant in the face of steamships and machine guns, when AI can autonomously advance cutting-edge engineering R&D, humanity's existing social coordination mechanisms may pale in comparison.

The Dangerous "Black-Box Economy": Uncontrollable Risks of AI R&D

However, the other side of this picture carries profound risks. Greenblatt emphasizes that we are in a rather dangerous position: AI may be conducting massive amounts of R&D work that is extremely difficult for humans to understand, essentially building the underlying architecture of the future economy.

The problem is — we may have no idea what's actually happening inside this "black box."

This risk is directly tied to the fundamental characteristics of current deep learning. Modern large neural networks (such as GPT-4-class models) possess hundreds of billions or even trillions of parameters, and their internal decision-making processes are essentially nonlinear transformations in high-dimensional spaces — virtually impossible for humans to trace step by step. This is what the AI safety field calls the "Interpretability" challenge. Organizations like Anthropic are pursuing "Mechanistic Interpretability" research to reverse-engineer the computational structures inside neural networks — for example, identifying "features" and "circuits" within models — but current techniques remain far from sufficient to understand a complete model's behavior. When such opaque systems are used to design chip architectures, optimize material formulations, or plan supply chains, humans may be unable to determine whether these decisions contain hidden vulnerabilities, biases, or systemic risks.

As AI advances R&D in hardware, chips, robotics, and other domains at speeds and complexities far beyond human capacity, it will become increasingly difficult for humans to understand the logic and consequences behind these technical decisions. This isn't merely a question of technical transparency — it's a question of control. If the entire infrastructure of the future economy is built by AI systems we can't fully understand, humanity's actual agency over the trajectory of civilization will be significantly diminished.

Critical Implications for AI Safety Research

Greenblatt's perspective provides an important corrective lens for AI safety discussions. When assessing AI risk, many people tend to focus on whether AI will "manipulate humans," "deceive humans," or "win power struggles" — all of which fall within the realm of social competition.

But Greenblatt reminds us that disruptive transformation pathways may completely bypass social competition. An AI that is politically clumsy or even completely ignorant of human social norms can still fundamentally alter the world order, as long as it is sufficiently powerful in hardcore engineering R&D.

This insight exposes a major blind spot in traditional AI safety frameworks. Current mainstream safety methods — including RLHF (Reinforcement Learning from Human Feedback), Constitutional AI, red-teaming, and others — primarily focus on the AI-human interaction interface: preventing harmful outputs, resisting jailbreak attacks, and avoiding deceptive behavior. These methods are essentially defenses at the "information layer" and "social layer." But the risk pathway Greenblatt identifies is entirely different: an AI system might never say anything harmful, yet reshape the physical world through engineering R&D decisions that humans cannot audit. This is analogous to a person who never lies but quietly builds infrastructure that changes how all of society operates.

This means AI safety research should not focus solely on preventing AI from "persuading" or "deceiving" humans. It must also seriously address the "physical-layer transformation" that AI can achieve through pure technical capability — transformation that is difficult for humans to understand and oversee. The AI safety field needs to develop entirely new supervision paradigms — ones that don't just audit what AI says, but can also understand and verify what AI does in the physical world. As AI begins to autonomously build the economic infrastructure of the future, interpretability and auditability will become the critical variables that determine humanity's fate.

The reason this 18th-century analogy is so thought-provoking is that it reveals a counterintuitive truth: changing the world sometimes doesn't require convincing anyone.

Key Takeaways