Explaining Language AI

Your manager asks you to explain how a chatbot works to the board. They've got no technical background, and five minutes. If you can do that well, then you'll be one of the most valuable people in the room.

• Explain it to the board
• Five minutes
• Be the most valuable person there

Explaining embeddings

Say, imagine a huge map, in which every word and sentence has a location, and things that mean similar things sit close together. Cat is near dog. A question about refunds is near the refund policy. Searching by meaning just means finding what's nearby. It's that simple, and that's why it's so useful.

Explaining an LLM

Say, it's like a very advanced autocomplete. It's read a vast amount of text, and it learned to predict what comes next. Do that word by word, and you get an answer. That also explains the problem. It's predicting what sounds right, and that's why it can invent a fact, in a very confident way.

Explaining RAG

Say, instead of asking someone to answer from memory, you give them the company handbook, open at the right page. First we find the right pages. Then we ask the model to answer from them, and to show where it found that. That's why it's much more reliable than the model on its own.

Summary