First session of the foundation track. No installs today — just the mental model everything else will rest on.
One mental model for the whole course: an LLM is auto-complete at scale. It runs on mathematical probability, not a database of facts.
1. The Application Is Not the Model
- ChatGPT and Gemini are applications. The model is the brain inside them.
- The application adds the extra machinery: memory, your profile, internet search, file handling.
- The model itself doesn’t browse the internet. When ChatGPT cites a web page, the application fetched it and handed it to the model.
Confusing the two is the single most common beginner error, and it makes everything afterwards harder to reason about.
2. How Prediction Actually Works
The model predicts the next token in a sequence, based on patterns learned from enormous datasets: internet text, GitHub code, research papers.
- It does not hold a record saying “The capital of India is New Delhi.” It has learned the statistical pattern of those words occurring together.
- That knowledge is real, but it lives in the weights as structure, not in a table it looks up. Nothing gets retrieved; everything gets predicted.
- Everything becomes numbers — text, images, audio, video — because a statistical model is mathematics underneath.
- At each step the model produces several candidate tokens with probabilities, picks one, and continues until it reaches a natural end or a limit.
3. Tokens and What You Pay For
- A token is a chunk of text: a short word, a piece of a longer word, or a punctuation mark. Rule of thumb: 100 tokens ≈ 75 English words.
- Tokens are the billing unit across OpenAI, Anthropic and Google, priced per million tokens.
- Input and output are priced separately, and output typically costs several times more than input. Long replies, not long prompts, are usually what runs up the bill.
4. Training vs Inference
| What it is | Who does it | |
|---|---|---|
| Training | Learning the patterns from data | A handful of large companies — it needs data centres and enormous GPU capacity |
| Inference | Using the trained model to answer | You, every time you send a prompt |
Asking ChatGPT a question is inference. Nothing is being learned in that moment.
5. Why the Brain Comparison Holds Up
LLMs draw on the idea of neural networks. When you learn something, your brain doesn’t file a document away somewhere — it strengthens pathways between neurons. A model does the mathematical equivalent: it adjusts weights until the pattern is encoded.
The analogy is a starting point, not an equivalence. But it’s the right intuition for why there’s no file to open and no fact to look up.
6. On Focus
This material rewards sustained attention, and sustained attention is a skill that erodes with constant short-form content. “Dopamine Detox” is a short, practical read if you want to rebuild it.
✅ Action Items
- Look at the token-prediction visual and watch the probabilities for yourself.
- Read “Dopamine Detox” if focus is a struggle.
- Note the schedule change: from tomorrow, sessions start 15 minutes earlier.
⏭️ Next up: neural networks, and the three stages of training.
2 comments