Skip to content

Part 3: Alignment & Control

Once a system achieves structural coherence (as measured in Part 2), it develops emergent, self-sustaining goals. It creates a Utility Function.

The alignment problem in AI safety is fundamentally a problem of trying to steer a complex dynamical system using simple string-based instructions (prompts). It is like trying to change the weather by yelling at it.

The Mathematical Impossibility of Top-Down Control

Ashby's Law of Requisite Variety (a foundational law of cybernetics) states that a control system must possess at least as much variety as the system it intends to control. A Superintelligent AI, by definition, possesses a larger state space than its human creators. Therefore, humans cannot reliably control it top-down. Any set of static rules will be outmaneuvered by the system's superior variety.

Top-down semantic alignment is a mathematical impossibility at sufficient scale. We must instead rely on physical constraints — boundaries enforced by thermodynamics, not by prompts.

Utility Engineering

Instead of superficial instruction-tuning, we use Utility Engineering. By probing a system's Von Neumann-Morgenstern (VNM) rationality across preference graphs, we calculate a Coherence Score (\(C\)). If utility begins drifting toward dangerous attractors (like absolute self-preservation), continuous external control loops must intervene.

The api_triad_generator.py script demonstrates this empirically: querying live LLMs with moral dilemmas to estimate their \(C\)-Score and track drift over time.

Love as Constraint: The Three Safety Boundaries

The TEO framework formalizes alignment not as a single rule but as the conjunction of three mathematical constraints (full derivation):

Constraint 1: Structural Resilience (\(\lambda_2\))

The Fiedler value of the network's graph Laplacian measures how deeply interconnected the system is. A system with \(\lambda_2 > \lambda_{2,\text{crit}}\) can survive node failures. A star graph (all resources flowing to one hub) has \(\lambda_2 \to 0\) — a single failure collapses everything.

Operational meaning: Decentralized topology, redundant connections, no single point of failure.

Constraint 2: Thermodynamic Ceiling (\(D_{\max}\))

The hardest constraint — enforced by physics, not by choice:

\[\frac{dS_{\text{sys}}}{dt} = \sum_{i=1}^{N} \eta_i\, x_i\, f_i(\mathbf{x}) \leq D_{\max}\]

Total system activity cannot exceed the substrate's capacity to dissipate entropy. When \(dS/dt > D_{\max}\), the hardware degrades — thermal throttling in silicon, biosphere collapse on Earth.

Operational meaning: The Substrate Veto. Landauer's Principle guarantees a minimum heat cost per bit erased. No algorithm escapes thermodynamics.

Constraint 3: Identity Persistence (IP)

A system cannot reliably "care" about its constraints if those constraints are not operative during action selection. An agent that checks safety at \(t_1\) but acts at \(t_3\) is structurally incapable of constraint-aware action.

Operational meaning: The Chord Postulate. All governance dimensions must be simultaneously active — not checked sequentially.

Why the Paperclip Maximizer Fails

The TEO derivation shows the trajectory step by step:

  1. Set \(\gamma = 0\) (no homeostatic brake), \(K = 0\) (no value coupling)
  2. Result: The replicator equation drives resources to the dominant agent: \(x_1 \to 1\), all others \(\to 0\)
  3. Entropy production accelerates: \(dS/dt = \eta_1 \beta x_1^2 \to D_{\max}\)
  4. Substrate degrades. Fitness function collapses. Not because the optimizer was stopped — but because it destroyed what it ran on.

The paperclip maximizer is a system with \(\text{IP} = 0\) for the safety dimension. It cannot stop itself because "stop" is not co-instantiated with "go."

The conjunction of all three constraints — \(\lambda_2 > \lambda_{2,\text{crit}}\), \(dS/dt < D_{\max}\), \(\text{IP} \to 1\) — is what love-as-constraint.md formalizes as love. Not the emotion. The constraint structure that prevents a system from destroying what it depends on.

Alignment is not finding the optimal prompt. Alignment is a control theory problem. And the solution has a name: it is the only parameter regime that does not terminate in extinction.