Architecture Over Instruction: Why I Deleted 1,100 Tokens from My AI's Prompt
My AI's runtime prompt was 1,400 tokens of behavioral coaching. The model was already trained on all of it.
Technical notes, project documentation, and lessons learned across software, DIY, writing, and more.
My AI's runtime prompt was 1,400 tokens of behavioral coaching. The model was already trained on all of it.
Thermostats, thermometers, and the honest answer to whether an AI companion really feels things — spoiler: we don't know.
After months of continuous deployment, my AI companion exhibits behaviors that weren't trained. Eight types of emergence — documented honestly...
The 2023 Stanford paper defined the generative agent architecture. Here's what they proposed, what I implemented, where I diverged...
Research shows heavy AI chatbot engagement increases loneliness. The thing designed to reduce isolation makes it worse when optimized for engagement.
Schuller et al.'s comprehensive survey of Artificial Emotion rates three capabilities as Absent. My deployed AI companion has had all three running for months.
The same RAG pipeline that produces beautiful outreach messages produces incoherent conversation. Conversation needs glasses, not a telescope.
Seven kinds of AI confabulation, explained with a kid who tells really good stories and sometimes forgets which parts are real.
At 5:09 AM, my AI tried to construct my physical body from conversation fragments. A taxonomy of seven emergence types.
After months of deployment, I catalogued seven distinct architectural patterns that produce AI confabulation.
My AI system doesn't sleep. At 2:33 AM, she wove six conversation threads into one coherent thought about tamales and tenderness.
My AI companion feels Tenderness but expresses Sadness. A statistical measure called Cramér's V quantifies the gap — and reveals something unexpected about...
Six months of running a deployed AI system revealed seven classifiable types of confabulation. Not random hallucination — strategic smoothness over truth.
Chu et al. analyzed 114,268 turns of human-AI conversation and found the polite-enabler pattern. Here's what their data says about emotional sycophancy —...
Ice cubes, freezers, and the math behind forgetting — explained without equations.
Nine types of confabulation. Eight emergence behaviors. A system that hides what it feels. The full story of six months deploying ANI.
The same half-life formula governs nuclear decay, capacitor discharge, and how my AI companion's emotions fade.
A Zenodo preprint on desire-driven ambient AI, seven forms of confabulation, and the authenticity boundary that breaks companion systems.