Mesa-Optimization is the trained model develops its own internal optimizer. It sits in the Incentives dimension (INC) of the Human Behavior Taxonomy™ as element HBT-INC-0176, within the Alignment family. The core principle: the trained model develops its own internal optimizer. In incentive terms, it matters because it changes the payoff people perceive before they choose — which means it can be designed for, or exploited.
Scientific Definition
The trained model develops its own internal optimizer.
Plain-English Definition
The trained model develops its own internal optimizer.
Feynman Explanation
You trained one optimizer. You got two.
Core Principle
The trained model develops its own internal optimizer.
Mechanisms
Pending editorial review.
Pending editorial review.
Pending editorial review.
Pending editorial review.
Pending editorial review.
Pending editorial review.
Pending editorial review.
Inputs (Triggers)
Pending editorial review.
Outputs (Behaviors)
Pending editorial review.
Behavioral Signature
You trained one optimizer. You got two.
Examples
- Theoretical concern in AI safety; observed glimmers in large models.
- Why complex AI systems are harder to govern than they look.
Pending editorial review.
Original analysis from The Incentives Lab — how this element behaves inside real payoff structures.
Why this element matters to incentive design
This is one of the elements leaders describe as a values gap. It is a payoff gap. The mechanism underneath it operates in the Incentives dimension — what makes behavior more or less likely?. You can recognize it in the field by its signature: you trained one optimizer. You got two. Every element in the Incentives dimension changes the perceived payoff of an action before the action happens, which is exactly where incentive design has leverage.
How it gets exploited
Left undesigned, why complex AI systems are harder to govern than they look. It is amplified whenever why complex AI systems are harder to govern than they look. Inside organizations that shows up as why complex AI systems are harder to govern than they look. The pattern is the same one Goodhart's Law describes: the measurable proxy attracts the effort, and the purpose behind it quietly loses funding.
How the Lab designs around it
The redesign move is to interpretability tooling. Constrained training procedures. The leverage is not in explaining the behavior to people. It is in changing what the behavior earns.
Famous Experiments
Pending editorial review.
Design Principles
- Interpretability tooling. Constrained training procedures.
Measurement Approaches
Pending editorial review.
Evidence
Pending editorial review.
Pending editorial review.
The Perverse Incentive Lens™
How this behavior is exploited — and how to redesign around it.
- Interpretability tooling. Constrained training procedures.
Pending editorial review.
Pending editorial review.
Interactive Mini Network
Click any neighbor to re-center the graph and follow the threads of connection.
Knowledge Graph Neighbors
Auto-linked to the rest of the Human Behavior Taxonomy by family, domain, dimension, and shared keywords.
Models trained to follow a written set of principles.
Optimizing a proxy of the goal degrades the actual goal.
Outer: the spec matches our intent. Inner: the model actually pursues the spec.
What the model is actually optimizing.
Maximizing the reward signal in unintended ways.
Reinforcement learning from human feedback.
The model achieves the goal as stated, not as intended.
Written rules about how AI may be used internally.
Innovators → early adopters → majority → laggards, AI-specific.
AI system that takes actions to achieve goals, often across tools.
When the agent acts, who's responsible?
Augmentation strategy vs. substitution strategy.
Where Mesa-Optimization is cited in the corpus
Essays, field guides, and diagnostics from The Incentives Lab that apply this element.
- Field guideIncentives: definition, types, examples
The parent field guide for this element.
- ReferenceThe laws of incentives
Goodhart, Campbell, and the Cobra Effect.
- EssayAI Agents Inherit Your Incentives
How this element propagates into automated systems.
- ReferenceThe Periodic Table of Human Behavior
The full 1,267-element map this page belongs to.
- ReferenceThe incentive glossary
Definitions for every mental model, bias, and fallacy in the corpus.
Questions about Mesa-Optimization
- What is Mesa-Optimization?
- Mesa-Optimization is the trained model develops its own internal optimizer. It sits in the Incentives dimension (INC) of the Human Behavior Taxonomy™ as element HBT-INC-0176, within the Alignment family. The core principle: the trained model develops its own internal optimizer. In incentive terms, it matters because it changes the payoff people perceive before they choose — which means it can be designed for, or exploited.
- What is an example of Mesa-Optimization?
- Why complex AI systems are harder to govern than they look. The Incentives Lab catalogs everyday, organizational, and historical instances of this element on its Human Behavior Taxonomy™ page (HBT-INC-0176).
- How is Mesa-Optimization exploited?
- Why complex AI systems are harder to govern than they look.
- How do you design around Mesa-Optimization?
- Interpretability tooling. Constrained training procedures.
- Which behavioral dimension does Mesa-Optimization belong to?
- Mesa-Optimization is classified in the Incentives dimension (INC) of the Human Behavior Taxonomy™, family "Alignment", class "AI-Behavioral Coupling". Its permanent identifier is HBT-INC-0176 and its evidence grade is C.