Personality Architecture Engine
PAE
PAE is a personality generation model for deciding what an AI values and where it draws the line, then checking that those choices hold as situations change.
Judgment before voice.
In detail
Settings fade as conversations get longer
When an AI is given a personality, it is usually through a written profile: personality, backstory, a way of speaking. That works well for short exchanges. Over a long conversation, or under steady pressure from the other side, the original personality can start to wear thin.
The AI model supplies what an AI can do. What it cares about, and where it stops, is often left blank. That gap is where PAE works.
Judgment before voice
PAE starts with what the AI's personality wants, what it protects, and what it will not give ground on. The way it talks is left to follow from that.
The aim is for its words and actions to change with the situation while the way it decides stays the same.
PAE does not train a new language model. It sits on top of one, designing and checking the standards an AI's personality keeps.
How it works
Design
We set out each personality's judgment in a document people can read and revise.
Run
We carry those standards into every conversation for the AI model to follow.
Check
We test whether the standards hold over long conversations and difficult moments, fix what goes off, and check again.
Now and next
Its first application is Luen, our own AI conversation service, where we refine the approach.
Next, we intend to extend it to other fields where AI meets people, such as education, training, games and robotics.
Partnerships
If you are building a service or product where an AI has to remain one person for a long time, we would like to hear from you.