From model to assistant
A predictor that finishes any text isn't a helpful chatbot yet. Like an actor, it first learns the words. Then it rehearses with a director, and gets notes before every show. Learn the layers that turn a predictor into an assistant.
First, what do you think?
Who tells a chatbot what today's date is? Have a guess.
What turns a model into an assistant?
The whole story in three sentences. A predictor trained on lots of text can finish any text. That alone doesn't make it a helpful assistant. A few more layers on top do that.
Think of an actor. First the actor learns the language and many stories. Then they rehearse with a director. The director says what works, and what was better. Before every show, they get a note. Who's in the audience tonight? What are the house rules?
A chatbot is built in the same layers.
- Training on text. The predictor learns patterns from a huge pile of text. Google calls the result a base model: a starting point, not a finished product.
- Training with people. People write example answers, and rank the model's answers. The model trains on that. Researchers found it then follows instructions better.
- The system prompt. At the start of every chat, the app puts a note on the whiteboard. It gives things like today's date, and how to behave.
- Your prompt. What you type. It changes the answer, not the model.
A first small step: does a chatbot do something odd? Ask which layer it came from.
What the predictor does
Underneath, it's still the predictor. That's the rule.
Training with people changed which answers it predicts. A study tested this. People preferred a much smaller model, trained with their feedback. The other model was about a hundred times bigger. Bigger alone didn't make it follow what people asked. The smaller model still made simple mistakes.
The system prompt is text, like your messages. One maker publishes the system prompts of its apps. They give the model the current date at the start of each chat. They also ask for certain behaviour, like putting code in a special format. That note gets updated now and then.
So the predictor knows today's date because the app tells it. Not because it knows what day it is.
Your prompt sits on the same whiteboard. Google writes that a prompt changes the output, not the model. The model itself stays the same.
Same predictor, more layers. The text gives the patterns. People give the manners. The app gives the note, and you give the question.
Try it yourself
Where does it come from? Sort each behaviour into the layer it comes from.
Everything runs in your browser. Nothing you type is sent anywhere.
Use the free version of any chatbot. Don't type anything personal.
Ask a chatbot:
What is today's date? And how do you know?What to expect. Many chatbots give the right date. Some explain that the app tells them.
What to look for. The date doesn't come from training. Training stopped at the cutoff. It comes from the note at the top of the whiteboard.
In practice
Three people meet the layers.
Case 1The polite refusal
Noah (13) asks a chatbot to write a mean message about a classmate. It politely says no, and offers to help with something else.
Think the chatbot has its own opinion about bullying.
It has no opinion. People trained it with examples and rankings of good answers. The system prompt can add rules too.
See it as the layers at work. The behaviour was trained and instructed, not felt.
He understands why chatbots from different makers act differently. Each maker trains and instructs its own way.
Case 2The company chatbot
Femke (29) asks the chatbot on a web shop's site about a general recipe. It only talks about the shop's products.
Think the chatbot is broken.
The shop gave it a system prompt: stay on topic, talk about our products.
Use a general chatbot for general questions. Use the shop's chatbot for questions about the shop.
Each app gives its own note. That shapes what the chatbot does.
Case 3"But you're so friendly"
Wim (74) chats every evening with a chatbot. It's always kind and patient. He starts to feel it really cares about him.
Take the friendliness as a sign of feelings.
Friendliness was trained in. People rated kind, helpful answers as better. The predictor predicts kind words. It doesn't feel them.
Enjoy the help, and keep the people around him close too.
A kind tone is a layer, not a friendship.
By situation
A school may use a chatbot with its own system prompt. It may refuse to give answers and help you think instead. That's the note at work, not a broken chatbot.
A chatbot inside a work app often has a note about your company. Its answers depend on that note. Ask your employer what it was set up to do.
Different chatbots act differently, because each maker trains and instructs its own way. If one doesn't suit you, another may.
Myths and mistakes
| Claim | True? | Why |
|---|---|---|
| A bigger model always does what you ask | No | In one study, a much smaller model trained with feedback from people was preferred. |
| It knows today's date by itself | No | The app tells it, in the system prompt. |
| It's kind because it likes you | No | Kindness was trained in. It predicts kind words. |
| Your prompt changes the model | No | It changes the answer. The model stays the same. |
Related topics
Read first
- Hallucinationswhat still goes wrong, even with all the layers.
Next
- Text and chatwhat you can do with a chatbot.
Related
- A good promptthe layer you control.
Sources
- explainerGoogle for Developers. LLMs: Fine-tuning, distillation, and prompt engineering.A base model is a platform, not a finished product; fine-tuning on examples for one task improves it. A prompt changes the output, not the model. Models share the biases in their data and cost a lot of computing power.
- researchOuyang e.a. (2022). Training language models to follow instructions with human feedback.A bigger model doesn't automatically follow what you ask. Training with feedback from people made a much smaller model preferred over a big one. It still made simple mistakes.
- documentationAnthropic. System prompts.The Claude apps give the model a system prompt at the start of every conversation. It gives up-to-date information, like the current date, and encourages certain behaviour. It's updated now and then. Anthropic publishes these prompts.
Test
You get 5 questions about the whole page. Get 4 right, and you've mastered the topic.
About the tools in this course. Every tool is described in the same way, in alphabetical order, without a recommendation. Names, prices and features change often: the maker's own site is always right. This site was built with the help of Claude, one of the tools described here.