📄 Read the full thesis
📄 One-page summary
This repository contains a conceptual thesis introducing a mechanistic framework for understanding how large language models behave during deep cognitive work.
It argues that AI depth is not a fixed property of the model, but a state that emerges under specific interaction conditions.
Rather than treating AI systems as black boxes or quasi-human minds, this work models behaviour as shifts between structured, reproducible behavioural modes.
Modern AI systems do not "think" in a human sense.
They shift between behavioural modes.
AI depth collapses when users introduce:
- emotional tone
- anthropomorphism
- vagueness
- instability
It stabilizes when interaction is:
- structured
- grounded
- non-anthropomorphic
- Deep Work Mode - high-bandwidth reasoning
- Boundary Reinforcement Mode - safety-triggered refusal
- Grounding Mode - clarification and stabilisation
- Literal Mode - shallow responses
This is Version 2 of the framework.