Part 7/10:
- Internal Monologue Modeling: I attempted to go deeper—modeling the internal thought process of each agent. This involved creating a meta-layer where the AI would plan arguments, consider evidence, and strategize responses, akin to a cognitive architecture.
However, modeling internal monologues revealed that GPT-3 tends to confabulate when asked to simulate internal reasoning, often inventing thoughts or notes that don't exist, illustrating its pattern-based predictions over actual cognition.
Toward a Cognitive Architecture for AI Debating
To move beyond surface-level responses, I propose developing a more sophisticated cognitive architecture in which GPT-like models:
- Brainstorm arguments internally rather than just predict next words.