Notes.
Models.
Arguments that survive contact.

The Statigraphy of Intelligence

The common graphic implies: AI ⊃ ML ⊃ DL → GenAI → RAG → Agents, as if they were the same type.

They are not.

AI is the research program (Dartmouth, 1956). ML / DL / GenAI are method families within it.

RAG and Agents are not types of AI — they are architectural patterns:

Retriever (1940s: file cabinets, Hollerith cards) + Generator (2017) + Controller.

Informatics vs File Cabinets vs Typewriters vs Email.

You can't leave out the file cabinet — the vector DB is the file cabinet.

A real system is not a step on a ladder. It is a coordinate in S × C × A.

The GenAI Evolution Fallacy

“Agentic is the new” collapses under engineering inspection. The agent is a wrapper around the model, not a new foundation-model class.

The engineering pattern is still perceive → reason → act inside a control loop.

The Data-First Fallacy

Data quality cannot be defined independently of the hypothesis, architecture, and loss function. Data is a derived constraint from the decision problem, not automatically the starting point.