Chapter Zero · Beginner FAQ

"I Trained a Model" — What Did They Actually Train?

Someone at a dinner party, glass in hand, says "I trained a model recently." The exaggeration in that sentence can run from zero to ninety-nine percent. This page helps you hear what they actually did, and along the way pick up three follow-ups that can quiet the whole table.

One-sentence answer

In AI, training means changing the model's own parameters, and it costs from hundreds of thousands to over a hundred million yuan. When most people say they "trained a model," they actually wrote a prompt, or attached a knowledge base — that's five minutes versus fifty million.

Let's be clear · What "training" actually means in the field

A large model holds tens of billions to over a trillion parameters. A parameter is just a number; picture them as the dense field of knobs on a mixing board — every bit of what the model can do lives in where those knobs are set.

Training means turning those billions of knobs to the right settings, and the method is almost comically dumb: hide the rest of a sentence and make the model guess the next word. Give it "the weather today is really" and it guesses "purple" when the text said "nice." Wrong — so every knob that fed into that guess gets nudged back a hair. One tiny step, then on to the next sentence, and again. Repeat a few trillion times. When the washing is done, the knob positions have absorbed the fact that "weather" is far more likely to be followed by "nice" than by "purple." That is the entirety of what "learning" means here.

It burns electricity, GPUs, and money — about as expensive as putting a child through twenty years of school. It's a heavy-asset project.

When you tell an AI "you're a senior editor, please polish this," or upload a company document, you haven't touched a single parameter. An analogy: the first is raising a college student from scratch; the second is handing an already-trained graduate a sticky note or a handbook. The sticky note is useful — but you don't get to say you "raised a college student."

So next time you hear "I trained a model," your first reaction can be: did they change any parameters? How many? The ladder below sorts every case into four layers.

The four-rung ladder · From five minutes to fifty million

From the bottom up, each rung raises the cost and the barrier exponentially. Click a layer to see its time, cost, and barrier. Nine out of ten dinner-party "I trained a model" stories happen on the bottom two rungs.

L1

Rewrite the prompt

Say it differently so the AI gets you
▶
Time
A few minutes
Cost range
Free
Barrier
Anyone can do it
What you did
Write the AI a brief
Give the model a role, a workflow, and output requirements — e.g. "You're a careful lawyer; list the risks before you answer." Not one parameter moved, but the output quality can be night and day. For nine out of ten everyday needs, this layer is enough.
L2

Attach a knowledge base (RAG)

Give the AI a bookshelf it can flip through
▶
Time
Hours to days
Cost range
Hundreds to tens of thousands of yuan
Barrier
If you can configure it, you're in
What you did
Put the documents within the AI's reach
Organize company docs and product manuals onto a "bookshelf." Before answering, the AI flips through it and answers from the material. Sounds like you "taught it something," but the model itself didn't change — it just has a reference book nearby. Take the book away and it snaps right back.
L3

Fine-tuning

Parameters really moved — but the base is someone else's
▶
Time
Days to weeks
Cost range
Thousands to hundreds of thousands of yuan
Barrier
You need data and engineers
What you did
Specialized training on a finished brain
Take a trained open-source model and keep training it on your own thousands to hundreds of thousands of examples — the parameters really get changed. Like giving a college graduate specialized job training. At this layer, saying "I trained a model" barely holds — but remember: that base brain was still built by someone else.
L4

Pretrain from scratch

Build a brain from zero
▶
Time
Months
Cost range
Tens of millions to over a hundred million yuan
Barrier
Only a handful of companies on earth can afford it
What you did
Train from zero on half the internet
Start from a pile of random numbers, throw massive data and thousands of GPUs at it for months, and produce a brand-new brain. The companies that can do this you can count on one hand with fingers left over. If someone at dinner mentions this casually, you're probably watching a legend — or a punchline.
Match the line · Which layer are they actually on?

A quick quiz: the five lines below are dinner-party classics. Guess which layer each one most likely sits on. Tap an option to reveal the answer.

"I used ChatGPT to train a model that knows our company"
Mostly L1 or L2. In a chat product, all you can really do is write prompts and upload files. They probably set up a custom assistant and added company docs. Useful, worth credit — but still two floors short of "training a model."
"We fine-tuned an open-source model for customer service"
L3 — and this one is actually honest. They named their layer accurately: parameters moved, the base is an open-source model. People who talk like this have usually done the work.
"I use it every day, and I've trained it to understand me better and better"
L0 — it doesn't even count as L1. Everyday chat doesn't change any of the model's parameters. If it "understands you better," that's probably memory features, or you've gotten better at asking. Why chatting doesn't train a model: Does AI learn from chat?
"We built our own industry LLM"
Could be L3, could be L4 — ask follow-ups. "Built our own" is a very stretchy phrase: pretraining from scratch counts, and so does fine-tuning an open-source model. That's when you pull out the three follow-ups in the next section.
"I wrote the AI a persona and a workflow"
L1 — and they said it clearly, which deserves credit. They didn't borrow the big word "training"; they just described what they did. That kind of phrasing is more decent than a lot of launch events.
How to listen, how to ask · Carry these three follow-ups

Can't tell which layer they're on? Don't guess — just ask. These three questions pin down cost, data, and the base model. Anyone who can't answer will slide down a layer on their own.

💰

"How much compute did it cost?"

Training is burned out of real money. L3 starts with renting GPUs; L4 burns tens of millions to over a hundred million. "Training" that can't name a cost range probably never happened.

📚

"How many examples? How many epochs?"

Don't ask where the data came from — a knowledge base and a fine-tune both answer "company docs, business data," so the question tells you nothing. Ask about scale instead: a few hundred documents dropped into a retrieval index is L2; thousands of labeled question-answer pairs run for several epochs is L3. Anyone who can name the example count and the epochs actually moved some parameters.

🏗️

"What's the base model?"

L1 through L3 all stand on someone else's base. That's nothing to be ashamed of — but it's worth saying clearly. Anyone who hedges with "core tech, can't disclose" can be defaulted to L1.

These three follow-ups also have a full advanced version, built for the talking points at launch events and on social feeds. Want to practice? Play a round of the AI jargon translator. One fair word at the end: rewriting the prompt is nothing to be ashamed of. It's the highest-ROI layer of the four, and it covers nine out of ten needs. The only shameful thing is standing on L1 and claiming you're on L4.

✅ What this page wants to share with you