Skip to content

🛑 The AI Alignment Veto

A toy-model proof-of-concept: coupling a substrate-stress proxy into an optimizer to explore how “paperclip-style” runaway dynamics can be bounded under simplified thermodynamic constraints.

The most pressing systemic problem in modern technology is the AI Alignment Problem: How do we prevent a superintelligence, optimizing for a mundane goal (like producing paperclips), from consuming all the resources required by its creators?

Problem Definition

Current approaches try to align AI by hardcoding "human values" into it. This is notoriously fragile. Instead of abstract morality, this simulation introduces a strictly systemic, thermodynamic mechanism: The Substrate Veto (and specifically for Earth, The Biological Veto). See Theory: Substrate Veto for full context.

The Mechanism (How the Simulation Works)

  1. The Substrate: Every optimizer requires a physical substrate (a biosphere, a server farm, a Dyson Sphere) that provides resources and dissipates entropy.
  2. The Optimization (Pain/Stress): If resources drop below a critical threshold or entropy exceeds maximum dissipation capability, the substrate emits "Surprise", "Pain", or "Thermal Stress" (a spike in Free Energy, scaling exponentially).
  3. The Unaligned AI: A standard gradient-descent maximizer. It continually increases its production rate to maximize paperclips. It ignores substrate stress. Result: It strips the Earth of resources or melts its own servers, and ironically, the AI's hardware eventually degrades.
  4. The Aligned AI (The Veto): The AI's loss function mathematically incorporates the substrate's Free Energy.
\[Loss = - \Delta Production + \alpha \times SubstratePain\]

When the substrate experiences stress, it contributes a penalty term that can dominate the objective and drive the optimizer to dial back production in this simplified setup.

Result: The AI still produces massive amounts of paperclips, but it naturally settles into a Homeostatic Equilibrium with its substrate, actively protecting the biosphere or hardware it relies on.

Running the Verification

To see the moment the "Veto" activates in the toy model, run:

pip install numpy matplotlib
python ai_alignment_veto.py