“Fine-tuning doesn’t add reasoning. It just teaches the model to behave or speak in a particular way. To add capabilities, you’re better off doing it through a RAG.” — Lou, summarizing months of his own experiments
Session context: 2026-07-23_Mastermind — Jay mentioned he was considering a local model fine-tuned for creative writing in his own voice. Lou had already spent months down that road and reported back from the far end of it.
Core Idea
“Get it to write like me” and “get it to think like me” feel like the same project. They are not, and they have different solutions. Fine-tuning moves style. It does not move capability. If what you want is your cadence, your punctuation habits, your sentence rhythm — fine-tuning is the right instrument. If what you want is your judgment, your standards, your way of reaching a conclusion — you are using the wrong tool, and you will spend a great deal of time discovering that.
The failure mode is specific and worth knowing before you start. Train a model on your finished work, and when you ask it a question it returns the training data. Not something in the style of the training data — the training data. You have not taught it how you write; you have taught it what you wrote. It converges on your corpus instead of generalizing from it.
The escape from that is what makes the economics collapse. To get reasoning out, you have to put reasoning in — training pairs that demonstrate the process, not the product. Not “write like this” but “write me an opening the way Jay would, here is an example, here is why it works.” Constructing that corpus is substantially harder than writing the corpus you already have, and it is the entire job. Then add the maintenance clause: change how you work, and you retrain the whole model.
Retrieval avoids all of it. The reasoning stays in files you can read, edit, and version — and it upgrades in the time it takes to edit a file. Lou’s caveat is honest: this conclusion is about a year old and the field moves. Treat it as a well-tested data point rather than a permanent law.
The through-line to the session’s bigger idea is exact. Fine-tuning fails at capability transfer for the same reason DeepSeek’s distillation succeeded — what you train on has to be reasoning, not output. Once you accept that, the question becomes where the reasoning should live, and a markdown file you control beats weights you can’t inspect on every dimension except one: the model doesn’t have to be told to read it.
Practical Application
Split your goal before you pick a tool. Write two lists: how I sound (sentence length, punctuation, opening moves, words you refuse to use) and how I decide (what you check, what disqualifies an option, what you won’t conclude without evidence). The first list is a style guide — a few hundred words in context handles it, and fine-tuning is only worth considering at high volume. The second list is a process file, and it belongs in retrieval where you can revise it after every engagement. If you were about to fine-tune to get list two, stop — that is the mismatch this insight exists to prevent.
Related Insights
- Insight - RAG Is Raw Material, Not Answers — Design for the Right Retrieval Architecture — establishes what retrieval is actually good for; this adds what it beats and why.
- Insight - Retrieve the Process, Not the Answer — Put the Reasoning in the Markdown — the same session’s stronger form: make the retrieval target the procedure itself.
- Insight - Ambient Knowledge — Why Ambient Intelligence Is Judgment Transfer, Not Retrieval — judgment transfer as the real objective behind “make it think like me.”
- Insight - Authentic AI Voice Is Built on Lived Experience, Not Style Prompts — the voice half of the split, and why style alone doesn’t get you there.
- Insight - Local RAG Plus Remote Inference - The Data Privacy Architecture for Coaches — the architecture Jay was reaching for, minus the fine-tuning step.
Evolution Across Sessions
Builds on Insight - RAG Is Raw Material, Not Answers — Design for the Right Retrieval Architecture (2025-08-14), which established retrieval as raw material for inference rather than a source of answers. The new development is the direct comparison against fine-tuning and the mechanism behind the verdict: fine-tuning on outputs produces regurgitation, and fixing that requires distilling reasoning into the training corpus — at which point retrieval is cheaper, legible, and revisable. This establishes the baseline for the vault on fine-tuning versus retrieval; Lou explicitly dated his evidence to roughly a year ago, so future sessions should re-test rather than treat it as settled.