Borrowed Brains
A mid-size legal-services firm wants “our own AI, trained on our 40,000 contracts.”
Vendor A quotes a custom-trained model: six months, a serious invoice, “your data becomes your moat.”
Vendor B says something that sounds like a shrug: “We won't train anything. We'll use the same frontier model everyone can rent, and just… hand it the right contracts at the right moment.”
In the bake-off, Vendor B's untrained system answers better, cites its sources, and ships in three weeks. The managing partner is confused, and a little offended: how can no training beat training?
Put yourself in the room. You have a real budget and 90 days. What are you actually paying for when you 'train an AI on your data'?
Your call — $50k, 90 days, 40,000 contracts. Where do you put the money?
Pick one — committing first is what makes the answer stick.
the lesson continues after you choose
The obvious answer is B — fine-tune on our data. It made sense in 2023, when context windows were tiny, frontier APIs were pricey, and 'domain AI' meant baking your domain into the weights.
But it misses what fine-tuning actually changes — and what happened to the economics since. Three numbers moved, and together they flipped the default.
Start with the deep idea, which the original DemystifAI book taught with a guitarist: a guitar player picks up the ukulele far faster than a total beginner, because most of the skill transfers. That's transfer learning, and it's the founding bargain of modern AI: models arrive pre-trained on trillions of words at someone else's expense, and your job is only ever the last mile — adapting a borrowed brain to your task.
The ladder, bottom to top. Prompting: instructions and examples in the request — free, instant, surprisingly far-reaching; always exhaust this first. Retrieval (RAG): search your documents and hand the model the relevant passages per question — this is where knowledge belongs, because it stays fresh, respects permissions and can cite sources. Fine-tuning: further training on your examples — this is where behaviour belongs: house style, strict output formats, a narrow task done millions of times, or distilling a big model's skill into a small cheap one. Pretraining: building the brain itself — a handful of labs on Earth.
Why 2023 advice aged badly, in three moves: context windows grew from ~4,000 tokens to hundreds of thousands or more — you can now hand a model entire contracts, not snippets. Inference prices collapsed — Stanford's AI Index measured a ~280× fall in the cost of GPT-3.5-level performance in two years. And frontier models got so capable that a rented brain plus your documents beats a custom-trained lesser brain on most business tasks. Fine-tuning didn't die — it moved to a specialist role: form, not facts.
So Vendor B wasn't cutting corners, and the managing partner's offence was aimed at the wrong target. “No training” actually meant: borrow a brain that cost someone else nine figures, and spend your budget on the one thing they can't sell anyone else — getting your documents to the model at the right moment. The training was already done. The moat was never the model; it's the data, the evals, and the workflow.
Your rule: name the gap before you spend. Ask “is the model missing knowledge, missing behaviour, or just missing instructions?” Knowledge → retrieval. Behaviour, needed identically at scale → fine-tune. Instructions → prompt, and prototype it this afternoon. Climb the ladder only when the rung below demonstrably fails.