How LLMs Actually Work
Swipe to show menu
You've talked to Claude. Now let's open the hood.
From Words to Numbers
Your message becomes tokens. Computers can't read words, only numbers, so your text gets split into small pieces called tokens (roughly 4 characters each: a full word, part of a word, or punctuation), and each token becomes a number.
750 words ≈ 1,000 tokens. More tokens = higher cost, slower reply.
Predicting, Not Knowing
AI doesn't understand your sentence — it guesses the next token, one at a time.
It picks the top guess, then repeats for the next token, and the next — until it stops.
Where the Guesses Come From
AI learned these patterns from huge amounts of text, books, articles, code, forums in one big training round, not live and ongoing.
- Continuous, unchecked updates can make a model worse, not better, so training happens in batches, then gets released as a snapshot;
- That's why AI has a training cutoff: no idea of anything after that date, and no visibility into your company's data;
- It also learned the internet's mistakes along with everything else.
When It Gets It Wrong
A hallucination is AI sounding confident about something that's just false. It's not lying — it's still just predicting, and the most likely-sounding answer isn't always the true one.
Not a dealbreaker — grounding Claude in real documents, search, or data (what you're practicing in this lab) cuts this way down. Your move: stay a little skeptical on specifics dates, numbers, links, and double-check when something feels off.
Thanks for your feedback!
Ask AI
Ask AI
Ask anything or try one of the suggested questions to begin our chat
How LLMs Actually Work
You've talked to Claude. Now let's open the hood.
From Words to Numbers
Your message becomes tokens. Computers can't read words, only numbers, so your text gets split into small pieces called tokens (roughly 4 characters each: a full word, part of a word, or punctuation), and each token becomes a number.
750 words ≈ 1,000 tokens. More tokens = higher cost, slower reply.
Predicting, Not Knowing
AI doesn't understand your sentence — it guesses the next token, one at a time.
It picks the top guess, then repeats for the next token, and the next — until it stops.
Where the Guesses Come From
AI learned these patterns from huge amounts of text, books, articles, code, forums in one big training round, not live and ongoing.
- Continuous, unchecked updates can make a model worse, not better, so training happens in batches, then gets released as a snapshot;
- That's why AI has a training cutoff: no idea of anything after that date, and no visibility into your company's data;
- It also learned the internet's mistakes along with everything else.
When It Gets It Wrong
A hallucination is AI sounding confident about something that's just false. It's not lying — it's still just predicting, and the most likely-sounding answer isn't always the true one.
Not a dealbreaker — grounding Claude in real documents, search, or data (what you're practicing in this lab) cuts this way down. Your move: stay a little skeptical on specifics dates, numbers, links, and double-check when something feels off.
Thanks for your feedback!
