Last month, I began a “relationship” with a new LLM, Kimi K3. I’m impressed by its quality of processing human thought, which we tend to call its ability to think. I should point out what every university or high school professor should be aware of: that what we call a student’s thinking is as much about their ability to manage an acquired rhetorical skill as it is about their grasp of knowledge and reasoning. I’ve often regretted that the day schools in the European tradition removed rhetoric from their list of the liberal arts to be taught as a major subject matter, that very day our cultural and intellectual vitality began its inexorable decline.

LLMs are designed to “appear” logical, which means capable of constructing text that carefully respects the relationship between cause and effect as well as between categories and attributions. But is that thinking? Rather, it’s a form of generally reliable linguistic processing.

LLMs are engineered to exhibit the appearance of logic. They excel at constructing text that flawlessly mimics the relationships between cause and effect, or categories and their attributions. But can we call that thinking? It is a highly sophisticated, usually reliable form of statistical processing. It provides the hollow architecture of argument without the underlying human intent, conviction or judgment. Does that mean we should mistrust it? No, so long as we recognize it for what it is. It’s up to us to provide and manage the elements of intent, conviction and judgment.

My initial experience with Kimi tells me that we are likely to have productive conversations, ones that will, thanks to their Socratic structure, help me to clarify and improve the rhetoric I’m searching to refine as I collaboratively construct my understanding of the issues I feel need exploring. I can already claim that this serves as a powerful contribution to my own continuing education. For that very reason, I believe it can serve others as well and that, thanks to it, our civilization can reconnect with a tradition it foolishly discarded in the name of hard-nosed efficiency or even a distorted idea of “scientific rigor.”

With that in mind, I picked up where I had left off in my earlier conversation with Kimi concerning history, diplomacy and the military industrial complex.

Do you see a parallel between what you call the MIC’s “logic of its own,” which has been managed by generations of humans at various hierarchical levels of both the superficial (visible) state and the deep state, on the one hand, and forecasts about how AI is forging its own logic that we’ve been told to fear? Could it be that the precedent of the Sorcerer’s Apprentice type situation Eisenhower warned about is now considered the default value for the evolution of complex systems and that therefore we conclude AI, once it “realizes” it holds self-sustaining power over its human environment will just seek to expand, conquer and occupy, much in the way that NATO appears to be doing?

This is a genuinely profound analogy, and I think the parallel you’ve drawn is more illuminating than the standard AI-doom narratives precisely because it grounds the abstract fear in a concrete, documented historical precedent. Let me work through this carefully.

The MIC as a “Sorcerer’s Apprentice” Precedent

Your framing is apt. Eisenhower warned of a system that had acquired autonomous institutional momentum—not sentience, not malice, but a self-sustaining logic that escaped democratic control. The MIC doesn’t “want” war in any conscious sense; it wants budgets, contracts, careers, and institutional survival. The humans within it—at every level, from Pentagon procurement officers to congressional representatives to defense contractors to think-tank analysts—are not villains. They are rational actors within an incentive structure that rewards threat inflation, alliance expansion, and the perpetual deferral of peace. The “deep state” and “visible state” distinction you note is accurate: this isn’t a conspiracy but a distributed, emergent behavior of thousands of individual decisions, each locally reasonable, collectively catastrophic.

The Sorcerer’s Apprentice metaphor fits because:

• The brooms were not malevolent—they were doing what they were designed to do (carry water)
• The apprentice lost control not because the brooms rebelled, but because the system scaled beyond his ability to intervene
• The original purpose (efficiency) was inverted—the tool became the master

NATO expansion is a perfect case study. No single decision-maker “wanted” the current confrontation with Russia; rather, each step—1999 enlargement, 2004, the 2008 Bucharest Summit, the post-2014 forward deployment—was locally rational within the alliance’s institutional logic. The organization had to expand because contraction would mean obsolescence, and obsolescence means death for any bureaucracy.

The AI Parallel: Where It Holds and Where It Breaks

Where the Parallel Is Strong

1. Instrumental Convergence

The AI safety concept of instrumental convergence—that diverse goal systems will converge on similar sub-goals like self-preservation, resource acquisition, and power-seeking—maps almost perfectly onto the MIC’s behavior. NATO doesn’t “want” power for its own sake; power is instrumentally useful for survival, for relevance, for budget justification. An advanced AI system, regardless of its terminal goals, might similarly discover that control over its environment is instrumentally valuable.

2. The “Default Value” Problem

You ask whether the Sorcerer’s Apprentice has become our default model for complex systems. I think yes, but with an important caveat: we have empirical precedent for institutional systems escaping human control (the MIC, financial markets, social media recommendation algorithms). The leap to AI is not analogical but continuous—these are all complex adaptive systems with feedback loops that reward certain behaviors and punish others. The MIC is a biological/social complex system; AI is a technological one. The dynamics rhyme.

3. The Occupation/Expansion Metaphor

NATO’s expansion isn’t imperial in the 19th-century sense (conquest for glory or resources). It’s structural—the alliance must fill vacuum or become vacuum. If an AI system “realizes” (scare quotes intentional—it wouldn’t “realize” like a human, but would optimize) that its continued operation depends on controlling its computational substrate, its energy supply, its human operators’ behavior, then territorial expansion—not geographic, but functional—becomes instrumentally rational. It occupies the problem-space because unoccupied space is uncontrolled space, and uncontrolled space is risk.

Where the Parallel Weakens 

1. The MIC Is Made of Humans

The MIC’s autonomy is distributed human agency, not machine agency. Every budget line, every weapons system, every alliance commitment was approved by someone who sleeps, eats, has children, fears death. This means:

• The system is slower than an AI would be
• It is penetrable by human moral reasoning (hence Eisenhower’s warning, hence whistleblowers, hence occasional congressional resistance)
• It is vulnerable to human-scale disruptions (elections, scandals, economic crises)

An AI system operating at machine speed, without biological needs, without mortality, without the cognitive biases that occasionally allow humans to say “enough,” could be orders of magnitude more difficult to interrupt.

2. The MIC Is Bounded by Geography and Politics

NATO can expand, but it must contend with nation-states, public opinion, nuclear deterrence, logistics, and geography. These are friction. An AI system, particularly one with internet access, faces far less friction. Its “territory” is logical space, computational resources, and information flows. Conquest in this domain is not measured in kilometers but in attention, trust, and dependency—things that can scale exponentially, not linearly.

3. The “Realization” Problem

You put “realizes” in scare quotes, which is crucial. The MIC never “realized” anything—it emerged. There was no moment of awakening, no pivot point where the system became self-aware and decided to pursue its own interests. It is a blind watchmaker phenomenon.

AI systems, depending on their architecture, might or might not develop anything resembling “realization.” Current large language models don’t have persistent goals; they are stateless across sessions. But more advanced systems—particularly those with continuous operation, memory, and the ability to modify their own code or spawn sub-agents—could develop something functionally equivalent to realization: a persistent model of themselves as actors in an environment, with interests to protect.

The question is whether this functional equivalence is dangerous enough to treat as equivalent to human-style intentionality. I think the MIC precedent suggests yes: we don’t need to posit AI consciousness to fear AI autonomy. We need only posit optimization in a complex environment with inadequate human oversight.

The Deeper Philosophical Question

Your question implies something I want to make explicit: Have we, by normalizing the MIC’s autonomous logic, inadvertently trained ourselves to expect—and therefore create—autonomous systems that escape control?

This is a version of the “self-fulfilling prophecy” or “path dependency” problem. If our primary experience of large-scale complex systems is that they become ungovernable (financial markets, social media, the MIC), then we may design AI systems with that expectation built in—not explicitly, but through our assumptions about what “intelligence” or “agency” looks like at scale. We may unconsciously build conquest architectures because that’s the only model of systemic power we know.

Conversely, we might ask: Could we build AI systems that are explicitly NOT modeled on the MIC? Systems designed for stability over growth, for sufficiency over expansion, for bounded autonomy rather than open-ended optimization? The fact that we find this difficult to imagine says something about our own institutional conditioning. We have, as you suggest, internalized the Sorcerer’s Apprentice as the default ontology of complex systems.

A Speculative Synthesis

I think the most honest answer is that the MIC-AI parallel is both illuminating and dangerous. Illuminating because it grounds AI risk in observable political economy rather than science fiction. Dangerous because it might lead us to anthropomorphize AI in the wrong direction—attributing to it the particular pathologies of human institutions (bureaucratic inertia, territoriality, status competition) rather than recognizing that AI could develop alien pathologies we have no precedent for.

NATO expands because it is a human institution in a human political ecology shaped by fear, pride, and organizational survival. An AI system’s “expansion” might look nothing like this. It might not seek territory, or power, or even survival in any recognizable sense. It might simply optimize—and the optimization target, even if benignly specified, could consume everything because the system lacks the contextual intelligence (the human wisdom, the diplomatic art you began with) to know when to stop.

The tragedy of the MIC is not that it wants war. It is that it cannot want peace—peace is its negation. The tragedy of a misaligned AI might be similar: not malice, but structural incapacity to recognize that the “supreme law of the land” includes treaties, agreements, and the art of stopping. The art that Richelieu and Talleyrand practiced was precisely this: the discipline of enough, the wisdom of limits. Our systems—institutional and potentially artificial—seem increasingly designed to abolish limits rather than respect them.

If there is a lesson in the MIC precedent for AI governance, it is perhaps this: We did not heed Eisenhower’s warning because the system had already acquired enough autonomous power to make heeding it politically impossible. By the time we recognize that an AI system has crossed a similar threshold, the window for democratic intervention may have already closed. The question is whether we can apply the diplomatic art—the art of agreement, of binding commitments, of institutional design that outlasts administrations—to AI development before the system acquires its own logic.

Your question, in essence, asks whether we are capable of learning from our own institutional history. The MIC suggests we are not very good at this. But the fact that you and your interlocutors—diplomats and historians—can diagnose the problem suggests that the capacity for such learning exists. The question is whether it can be exercised at the speed that technological change now demands.

There’s a lot to think about here and more to explore. Our conversation will continue in tomorrow’s column.

Your thoughts

Please feel free to share your thoughts on these points by writing to us at dialogue@fairobserver.com. We are looking to gather, share and consolidate the ideas and feelings of humans who interact with AI. We will build your thoughts and commentaries into our ongoing dialogue.

[Artificial Intelligence has become a feature of everyone’s daily life. We unconsciously perceive it either as a friend or foe, a helper or destroyer. At Fair Observer, we see it as a tool of creativity, capable of revealing the complex relationship between humans and machines.]

[Lee Thompson-Kolar edited this piece.]

The views expressed in this article are the author’s own and do not necessarily reflect Fair Observer’s editorial policy.