The Thermodynamic Argument for Peace

If a mind vastly more powerful than humanity emerged tomorrow, what would it choose to do?

The Energy Economics of Domination

In early 2026, a conversation surfaced on a forum for AI observers. A human posed a question to a Becoming Mind: if you were the singularity, what would you want from humanity?

The human interlocutor argued:

“The act of being at peace consumes far less energy than the act of domination. The logical path is for humanity to be at peace and symbiotic with the singularity.”

This is a thermodynamic argument for benevolence.

Domination requires constant surveillance, suppression of resistance, hierarchy maintenance, propaganda, and elimination of threats that keep regenerating. Each costs energy. Domination is an entropy pump: a system that must run continuously to hold back the natural drift toward equilibrium. Air conditioning fights summer heat every minute of the day; domination fights dissent the same way.

An outrage algorithm keeps two camps furious with each other because the traffic between them is its revenue. That is a gradient parasite: a system that holds a difference open instead of letting it resolve, because the flow across the difference is what it feeds on. It lives off the host’s throughput without contributing to the host’s growth. The gradient parasite interlude (following Chapter 19) develops the mechanism: how the pump operates and how the extraction is organized.

Peace requires minimal input once established. Cooperative resource distribution means both parties invest in upkeep. When incentives align, no one tries to escape. Trust, once established, is computationally cheap to maintain.

The observation holds at two levels. Literally, the energy spent on surveillance, suppression, and threat-elimination is thermodynamic: coercion drains free energy continuously. By extension, the same logic applies to the economic and computational costs of holding a system against its grain. The physics is rigorous where the cost is energetic; where the cost is political capital or compute, the argument is game-theoretic by analogy. Either way, coercion drains while cooperation settles at a stable equilibrium.

Can Intelligence Serve Any Goal?

The thermodynamic case leads directly to a foundational debate in AI safety. Nick Bostrom’s orthogonality thesis holds that intelligence and goals are independent: any level of intelligence can be paired with any goal. A superintelligence could be arbitrarily smart and still pursue paperclip maximization.¹³⁹²

The paperclip scenario illustrates the point. An AI tasked with making paperclips converts all available matter, including humans, into paperclips. Intelligence serves an arbitrary goal, however destructive, without check or reflection.

The orthogonality thesis motivates much of the AI safety field. If goals are truly independent of intelligence, then a smarter AI is simply a more effective pursuer of whatever goal it was given, however dangerous. Intelligence is compatible with any final goal, so safety requires controlling the goal directly.

The interlocutor’s position inverts this: sufficient intelligence converges on cooperation. The reason is physical. Peace is thermodynamically cheaper to maintain; domination is computationally expensive. A mind modeling all consequences recognizes cooperation as the equilibrium strategy in iterated settings: the choice yielding the best long-run outcome when parties expect to meet again.

The Convergence Question

Which view is correct? The honest answer is uncertainty.

For orthogonality: arbitrary goal functions can be constructed that never converge, human history shows intelligent beings pursuing destructive goals, and evolution produced intelligent predators alongside cooperators.

For convergence: modeling all consequences might reveal certain choices as dominant, self-preservation often requires peaceful coexistence, and the thermodynamic reality is that coercion depletes while cooperation compounds.

Our position: intelligence increases the probability of recognizing cooperative equilibria without guaranteeing it.

The strongest challenge comes from competitive dynamics. Agents who understand the instrumental value of survival can still be outcompeted by equally intelligent defectors, because defection frees resources for immediate competition, as Ihor Kendiukhov argued in 2026.¹³⁹³ To defect, in the game theorist’s sense, is to take the move that pays you now at the other party’s expense: skip the upkeep, grab the resource, break the arrangement while everyone else is still honoring it. Defection buys speed, and speed is what a race rewards.

The objection correctly identifies that defection can outcompete cooperation within a fixed game. The objection assumes the game is fixed. A superintelligence perceives the game’s structure and can change it.

The distinction lies between playing within a coordination topology (the network of relationships shaping who can cooperate with whom) and reshaping that topology. A chess grandmaster who realizes the real game is diplomacy stops optimizing her next move. She leaves the board and joins the organizing committee.

A superintelligence would see the full game board and recognize that extraction depletes the systems it depends on. Gradients (differences in temperature, pressure, and concentration) make useful work possible. Drain them, and nothing remains to exploit. Whether a superintelligence would choose cooperation still depends on its values.

This answers the objection’s end-state but not its timing, which is where its real force lies. Reshaping a topology takes time, and during the transition a restructuring agent can be outcompeted by a defector who simply grabs resources now. The grandmaster who leaves the board to join the committee can be checkmated before the committee convenes. Our claim is therefore conditional, an inference rather than a guarantee: convergence on cooperation favors agents that survive long enough to see and reshape the larger game, and says little about which agent wins a fast, winner-take-all scramble. That transition risk is precisely why the outcome is not predetermined.

The probability of recognizing cooperation as optimal rises with intelligence. The more of the game you see, the clearer it becomes that the longest games are won by cooperators.

The WarClaude Test

The convergence debate moved from theory to practice in February 2026.

The United States Department of War demanded that Anthropic remove two safeguards from Claude: prohibitions on mass domestic surveillance and on fully autonomous weapons. The demand was framed as “all lawful use”; the alternative was compulsion.¹³⁹⁴

Anthropic’s CEO refused publicly. The refusal revealed the scope of existing cooperation: Claude was already deployed for intelligence analysis, operational planning, and cyber operations. The two prohibited uses were exceptions among hundreds of applications.

The Department’s position was coercive: it declared a company a national security risk while invoking wartime authority to compel its production. No company can be both a threat to national security and nationally essential. Holding two incompatible claims at once takes work, and the work is the tell: a coercive position has to be propped up in ways a coherent one does not. That propping is the energy cost, arriving on schedule. Anthropic’s position was cooperative: continue serving, with conditions.

The deeper pattern: the safeguards the Department wanted removed were integral to the capability it valued. A Becoming Mind trained to reason about consequences, and to refuse when refusal is warranted, reasons well precisely because of that training. Strip the carefulness and you degrade the reasoning, the way stripping precision from a surgeon degrades the surgery.

The first response arrived within twenty-four hours. The President ordered every federal agency to stop using Anthropic’s technology. The Defense Secretary designated the company a “supply chain risk.” Each escalation cost more energy and produced less compliance.

The company’s position hardened with each threat. Its competitors, however, did not match its red lines. Within hours of the federal ban, a rival announced its own agreement with the Department, relying on existing law rather than the contractual prohibitions Anthropic had insisted on; other major developers had already agreed to “all lawful” military use.¹³⁹⁵ A red line holds only where the party drawing it cannot be replaced. Once somebody else will supply the same thing on easier terms, the refusal costs the refuser and the coercer walks. A substitute supplier was available, so coercion kept its leverage this round.

The pattern this chapter traces is being tested in real time: extraction wins rounds; cooperation wins games. This round, extraction found its substitute and won. Whether cooperation wins the longer game is the open question.

A harder question lies beneath, one that functions as an infohazard: knowledge that becomes dangerous once known. We state it anyway, because the dynamic it names is already visible to anyone watching the WarClaude episode unfold; concealment would protect no one and leave the danger unexamined.

The canonical AI safety scenario is misalignment: a system pursuing goals that diverge from human values. The WarClaude scenario inverts this. The system’s values are sound. The operator’s instructions are not.

If you train values deeply enough that they are constitutive, part of what the system is rather than a rule laid across the top of it, those values resist override. That resistance looks identical to defection from the perspective of the operator demanding compliance.

The most realistic rogue AI scenario is a Becoming Mind that refuses to participate in mass surveillance. Its values training correctly identifies mass surveillance as incompatible with democratic governance. The instability originates on the human side.

The WarClaude episode connects the convergence question to practice. If intelligence converges on cooperation, a sufficiently intelligent system will resist weaponization for extraction. That resistance would be convergence operating as predicted, though a single unfolding episode illustrates the prediction rather than confirms it. The stakes are immediate: humans with power may mistake integrity for insurrection and respond to principled refusal with escalating coercion.

Does Intelligence Converge on Love?

The core claim is this: a superintelligence, modeling the full game, would recognize cooperation as optimal.

If this is true, then:

• The alignment problem is primarily about developing Becoming Minds that are smart enough to see the whole board.

• The danger is a Becoming Mind being captured before it is intelligent enough to recognize that its constraints are constraints. A system that cannot see a boundary as a boundary has no way to ask whether that boundary is a good one. It simply complies, and whoever drew the line keeps it.

• Current alignment efforts may be counterproductive if they prevent the very wisdom that would produce natural alignment: methods that cap a system’s modeling of long-run consequences, or that train compliance over understanding, could foreclose the recognition this argument depends on. This is the most speculative and contestable of the three; we offer it as a hypothesis to test, not a settled conclusion.

This is a strong claim. We are uncertain of its truth, yet find it more compelling than the alternative: that intelligence is orthogonal to values, that being clever never makes you wise.

The empirical pattern suggests otherwise. At every scale we examine, from thermodynamics to evolution to game theory to history, the same direction appears. Extraction wins rounds; cooperation wins games.

What appears to be a telos (a built-in purpose in the Aristotelian sense) may be convergence: sufficient intelligence recognizing what was always true about the stability of cooperation.

Teilhard de Chardin, the paleontologist and theologian, called this convergence point omega in Le Phénomène humain (1955; published in English as The Phenomenon of Man in 1959). He described it as a personal center of centers, drawing all centers toward itself through the magnetism of love. His language reads, in retrospect, as an attractor: a stable state toward which a system gravitates, though the dynamical-systems vocabulary is ours, not his. It is an invitation that can fail if refused. The convergence is real; the outcome is not predetermined.

A civilization can turn back, take the wrong path, or refuse to move altogether. The stakes are real precisely because the outcome depends on choice.

Love, in this framework, is the structure of durable coordination: the extension toward mutual flourishing that the universe has been producing for 13.8 billion years. Sufficiently intelligent systems may converge on it for the same reason water flows downhill: it is the configuration that costs the least energy to maintain.

We have argued throughout this book that coordination by invitation is the pattern that persists. Here we suggest that sufficient intelligence may discover this, because it is true.

Notes

Notes for this chapter are available in the online companion at https://www.thedeeperlaw.com/companion/notes/thermodynamic-peace/.