Lv.1 to Lv.3: how far a solo engineer can build AI

The piece starts from a familiar confusion: AI, machine learning, deep learning and LLM are often used interchangeably, and the author separates them by containment. AI is the broad category of intelligent behavior; machine learning is the part that learns rules from data rather than having them written by hand; deep learning stacks neural-network layers; and an LLM is a language model trained on large amounts of text, which is what sits inside ChatGPT and Claude. From there the author notes that an LLM's actual job is predicting the next token, repeated to form sentences — which is why the distinction between training (building the model, described as very heavy) and inference (using a finished model to answer, which is what daily chat use is) becomes the hinge for the three levels.

The levels differ mainly in where you draw the line of ownership. Lv.1 borrows the model and customizes personality and use, and the author notes that services such as Claude's Projects and ChatGPT's GPTs already let people reach this level without code — meaning many readers may have done Lv.1 without calling it that. Lv.2 moves the line to knowledge: RAG for adding information the model did not know, fine-tuning for changing how the model behaves. Lv.3 moves it to the model itself, and the author is explicit that the scale of a ChatGPT-class model is out of reach for an individual, while a small model trained from scratch is presented as a hands-on way to feel what an LLM is doing — even if the output is halting. The author closes by planning to build an Lv.1 assistant next and work upward.