AINoon Lesson 2

How generative AI works and how that informs how to best use it

The chatbot predicts what text comes next, one word at a time
The prompt is split into tokens (words or parts of words)
- Image and audio inputs can similarly be split into chunks to act as tokens
Tokens passed into an LLM, with each token represented as a number
Those numbers are processed by layers of “neurons”
The final layer outputs a number for each possible token representing the probability that token comes next
A token is randomly picked based on those probabilities
- Some chatbots let you choose a temperature - a lower temperature is less random
- i.e. The highest probability is more likely to be picked with a lower temperature
The chosen token is added to the prompt, and the whole process is repeated to get the next token
Each neuron multiplies each input by a weight, sum the result, and applies a transformation to produce a single output number
- Loosely inspired by a neuron in the brain
Millions of neurons are connected in a special pattern called a “transformer”
- The transformer architecture was the breakthrough that led to ChatGPT
- When someone says “OpenAI’s GPT-3 has 175 billion parameters”, they’re counting the weights
- Models keep getting bigger, which lets them more accurately predict which token comes next

Weights start as random numbers
Weights are updated by a process called pre-training:
- Uses millions of training examples from general sources, like the internet and books
- E.g. Leave-one-word-out examples:
  The Beatles were a ____ from England → band
- For each example, weights are updated slightly to get closer to the right answer - like tuning a dial
- This trains the LLM in language patterns AND general knowledge
Extra training based on human feedback makes chatbots respond better to questions and instructions
- E.g. Chatbot gives 5 answers and human picks the best

Generative	The model generates text (or other content)
Pre-trained	Pre-trained on many examples from general sources
Transformer	The kind of neural network architecture used

The generation process is not magic, just simple maths!
We know how AI works, but no-one truly knows why it works
- The billions of calculations across neurons are too complex
AI doesn’t “think” with logic - it’s predicting the best answer based on its training sources
- People say AI sometimes “hallucinates” wrong answers - but every response is generated the same way
- It generates answers that sound right - and it’s surprising so many actually are right!
It doesn’t train during your chats
- It “remembers” by including chat history in the prompt
- Even “memory” features just include past chats in the prompt
- Public chatbots may use your chats to train later models

Everything is a hallucination - but it’s correct surprisingly often
It might not always “pay attention” to everything in the prompt
It might not consistently give the same answer
Different models (or model versions) can give very different responses to the same prompt

Make your own chatbot to answer questions based on information in a document

Either you test your chatbot, or your users will!

Ask a list of questions you want it to answer
Amend instructions or sources to improve answers
Ask each question multiple times to check it consistently gives the right answer
Re-test all questions after any change
- Changes to instructions, knowledge, or model
- Using a system to automate testing helps!
Test edge cases - How does it respond to off-topic questions?

Make a list of questions you want your chatbot to be able to answer
- Feel free to use a different source document!
Ask each question and note whether it was correct
For questions it got wrong, modify the source to help it get the right answer
Test it again!
- Is it right more often?
- Is it wrong for any questions it got right before?

Extra: Learn more about how LLMs work with a brief explainer or a deeper dive (more maths) from 3Blue1Brown.