I put the LLM back, watched it hit a wall, and discovered that the real architecture maps to Peter Thiel's distinction: zero-to-one discovery layered on top of one-to-N execution, each tier handling only what the tier below can't.
I built an AI agent, watched it fail seven times, swapped the model, removed the model entirely, coded thirty rules in Python — and discovered that the question isn't whether to use AI, but whether the problem is actually probabilistic.
Why I spent weeks building a knowledge base before writing a single line of agent code — and why understanding the architecture first was the best decision of the project.
How a frustrating loop of muddled HealthKit code led to a local AI knowledge system, study material generator, and text-to-speech learning tool