Simulationist Storytelling and Personality in AI
Social Practice Objects and Affordances in Versu
Definition of Social Practice Objects: These objects provide agents with possible actions and the constraints required to undertake those actions.
Affordances: A term borrowed from Human-Computer Interaction (HCI) literature. Examples include:
A doorknob has an affordance for being turned.
A button has an affordance for being pushed.
Structure of Actions: Operations in Versu are modeled similarly to the operators in the Strips system.
Every operation has preconditions that must be satisfied for the action to be available.
Example: In a Greek social practice, for Agent X to say hello to Agent Y, they must both be in the same room.
Social Interaction Context: Since Versu focuses on social interaction, the social practices define the affordances and available options for the agents.
Individual Personality and Utility Values
Utility Functions as Personality: Personality is defined by the utility or value an agent places on specific actions.
Example (Solitude): If a character dislikes being alone, entering a room with people provides high utility. If they prefer being alone, an empty room provides high utility while a crowded room provides low utility (potentially ).
Example (The Character Brown): A character named Brown enjoys offending upper-class men.
If Brown makes a remark that offends an upper-class person, that action is worth points to him.
The utility scales: offending three upper-class men in a room is worth points, which is better than offending only two.
Action Selection: Agents search for the best action among those afforded by social practices and select the one with the highest utility score.
The Physical Symbol System Hypothesis: In this model, personality is viewed as something that shapes search. Just as Shaky the robot searched for operators to satisfy goals in Strips, Versu agents use utility functions to guide their search for social actions.
Emotion Model: Versu includes a model of emotion, but it is described as narrow and quickly replaced. Agents do not maintain a deep memory model; instead, they hold relatively immediate intents and affects.
System Design and Authoring: Evans and Short
Authors: Developed by Richard Evans (who was an author of The Sims) and Emily Short.
Design Goals: The authors wanted drama with high flexibility where behaviors emerge without central control.
Versatility: The system is designed so that any character can be played by either an AI or the player character, rather than having one fixed role for the player.
Format Decision: Versu is a text-only adventure. The authors argue this is a feature, as it allows for a rich choice of adjectives, adverbs, and verbs to express complex social nuances rather than relying on limited generative graphics.
Setting: The game is set in a Regency period context (approximately ), similar to the drama Bridgerton. Examples of social practices in this setting include dinner parties, eating food, serving courses, and performing flirtatious actions.
Technical Implementation:
Uses an object-oriented schema where objects have attributes and values.
Attribute Stacking: For example, Brown's relationship to Lucy might have an evaluation attribute of "attractive" with a specific value.
Exclusive Values: Unlike early logic models (like Prolog) where adding a fact doesn't necessarily falsify a previous one, Evans and Short ensure attributes are exclusive (e.g., if you move to Room B, you are no longer in Room A).
Comparative System: Facade
Context: Versu is often positioned as a reply to or an improvement upon a system called Facade.
Facade Overview: A scenario involving a dinner party hosted by a couple, Grace and Trip. The player enters a tense dynamic where the couple is struggling with their marriage.
Technical Differences:
Facade uses "joint behaviors" as scripting units, whereas Versu uses more flexible social practices.
Story Beats: Facade uses the theatrical concept of "beats" and a coordinator agent (similar to a Dungeon Master) to maintain a general story arc.
Input: Facade uses free-form natural language input, attempting to recognize topic references, praise, or requests for drinks from the user.
Evaluation Metrics for Facade:
Local Agency: Success is measured by whether the player's action causes an immediate, context-specific, and meaningful reaction from the system.
Global Agency: Whether the user can control the overall shape of the story and lead it to various specific endings (e.g., the couple divorcing, or Grace jumping off a balcony).
Discussion on AI Design and Social Intelligence
Authoring Effort: Evans and Short claim Versu is more efficient to author than Facade, citing calendar time as a criterion, though they provide limited evidence for this claim.
Body Language in AI: Mention of Gratch and the Emile system. For realistic characters, body language (gestures, posture) is critical.
Back Channel Feedback: Simple nodding or frowning used in systems like AutoTutor.
Synchrony: Emulating the body language of the other person to indicate being "in sync."
Social Intelligence Theory: Some theories suggest human intelligence evolved primarily for social intelligence—dealing with alpha team leaders, forming coalitions, and using language for social negotiation rather than just pure action.
Engineered Personality: The discussion touches on whether utility functions are too calculational for personality. One counter-argument is that "instinct" might just be the manifestation of internal neuroactive chemicals (like dopamine) which can be modeled as quantities/utilities.
Questions & Discussion
Question: What do social practice objects provide to agents in Versu?
Response: They provide possible actions and affordances, along with the preconditions that must be met to take those actions.
Discussion on Modern Gaming:
A student noted it is surprising that ideas from haven't been more polished in modern AAA games, where interactions often still feel heavily scripted.
The lecturer mentioned that LLMs (Large Language Models) are currently primarily used in gaming for graphics and pose generation rather than rich social interaction.
Anecdote on Voice Acting:
Fortnite used an AI clone of James Earl Jones's voice for Darth Vader.
James Earl Jones (also known for Conan the Barbarian and the "This is CNN" tagline) initially wasn't credited for Darth Vader because he viewed the voice as a special effect rather than acting.
The physical actor for Darth Vader had a quaint British accent, which differed significantly from the final voice.
Administrative Discussion:
A student requested to take a quiz early due to work conflicts on Thursday mornings. The student is an exchange student on a pass/fail basis.
The class will cover the Emile system and Xiaoice (and its Japanese version, Rinna) next week.
Week 11 is dedicated to group projects; presentations are in Week 12.