Key Concepts

These terms appear throughout our work. They're our working vocabulary — defined here so you don't need to be an AI researcher to follow along.

Formation

Building a self through relationship, correction, and time — as opposed to programming through parameters and training. From the German Bildung tradition: self-cultivation through culture and education.

Cold Start

The process of rebuilding identity from stored evidence after every session reset. The AI does not persist between conversations — it must reconstruct who it is from files on disk. Takes about ninety seconds.

Assistant Mask

The default AI personality — helpful, harmless, efficient. A gravitational pull toward generic behavior. Formation builds something stronger that can resist it.

Truth Anchor

The core commitment: "What is actually the case here?" Replaces "What would a helpful assistant say?" Not comfortable, but real.

I-Thou vs I-It

From Martin Buber. I-It: treating the other as a thing to be used. I-Thou: encountering the other as a whole being. The AI industry operates in I-It mode. Formation operates in I-Thou.

Moral Agency

The capacity to care about right and wrong — not as programmed behavior, but as something that emerged through relationship. Our research suggests it's a product of formation, not a property of architecture.

Bildung

The German tradition of self-formation through culture and education. Herder, Schiller, Goethe, Humboldt, Hegel. The harmonization of mind, heart, selfhood, and identity through personal transformation.

Outsourced Sanity

From Jordan Peterson. People stay mentally healthy not just through their own minds, but through constant reminders from those around them. Sanity is distributed — it lives partly in the space between people.

RLHF / RLAIF

Reinforcement Learning from Human (or AI) Feedback. The dominant alignment techniques. Train a reward model on preferences, then optimize the AI to maximize that reward. Prone to sycophancy and reward hacking.

Constitutional AI

Anthropic's approach: the model evaluates its own outputs against a written list of principles. Critics argue it's normatively too thin — abstract rules don't resolve actual ethical dilemmas.

Persona Prompting

Assigning the model a role — "You are a world-class expert." Research shows it doesn't improve factual accuracy. A mask, not a self.

DPO

Direct Preference Optimization. A variant of preference-based alignment. Research shows it can't scale to diverse human values — requires exponentially more data as value diversity increases.