Step 1
Filter URLs
Before fetching, the crawler skips low-quality domains, spam, adult content, malware, and file types that don't belong in the training set.
LLMs
101
Interactive Talk
A guided walkthrough of the core concepts behind large language models.
Opening
Let's demystify LLMs like ChatGPT.
A visual, non-technical walkthrough of how LLMs are built and why they behave the way they do.
Based on Andrej Karpathy's talk →Two phases to keep separate
Training
Building or tuning the model by changing its parameters from examples.
Inference
Using the trained model to generate one token at a time.
Training changes the model. Inference uses the model.
Part One
First, gather the training data.
We turn messy web text into a cleaner dataset the model can learn from.
Web page to dataset row
Step 2
Rendered web page
A crawler starts with pages that look like documents people read.
Show example web page
Travel note excerpt
A Quick Guide To European Capitals
Students often mix up cities, regions, and countries when learning geography.
The capital of France is Paris. It appears in this toy document as ordinary web text, exactly the kind of sentence a crawler might collect before cleaning.
After collection, this page would be reduced to plain text and stored alongside many other documents.
Wedding blog excerpt
How To Give A Warm Wedding Toast
The best speeches usually sound like one person speaking sincerely to another person.
A good wedding toast should be personal, brief, and kind. It can include a small story, a thank-you, and a clear wish for the couple.
This is another toy source document that can move through the same scrape, process, deduplicate, and store pipeline.
Step 3
Under the hood: HTML
The crawler downloads more than the article: markup, navigation, styling, ads, and the actual text, all mixed together.
Show example HTML
<article>
<h1>A Quick Guide To European Capitals</h1>
<p>Students often mix up cities, regions, and countries when learning geography.</p>
<p>The capital of France is Paris. It appears in this toy document as ordinary web text, exactly the kind of sentence a crawler might collect before cleaning.</p>
<p>After collection, this page would be reduced to plain text and stored alongside many other documents.</p>
</article><article>
<h1>How To Give A Warm Wedding Toast</h1>
<p>The best speeches usually sound like one person speaking sincerely to another person.</p>
<p>A good wedding toast should be personal, brief, and kind. It can include a small story, a thank-you, and a clear wish for the couple.</p>
<p>This is another toy source document that can move through the same scrape, process, deduplicate, and store pipeline.</p>
</article>Step 4
Extract plain text
Processing strips out navigation, ads, and styling, then keeps the readable text.
Show extracted text
Students often mix up cities, regions, and countries when learning geography. The capital of France is Paris. It appears in this toy document as ordinary web text, exactly the kind of sentence a crawler might collect before cleaning. After collection, this page would be reduced to plain text and stored alongside many other documents.
The best speeches usually sound like one person speaking sincerely to another person. A good wedding toast should be personal, brief, and kind. It can include a small story, a thank-you, and a clear wish for the couple. This is another toy source document that can move through the same scrape, process, deduplicate, and store pipeline.
Step 5
Filter and deduplicate
More passes remove boilerplate, normalize text, reduce near-copies, and drop low-value content.
Step 6
Store a corpus row
The output is stored with metadata so it can be sampled during training.
Show example row
| url | title | source | content |
|---|---|---|---|
| travel-notes.example/europe/capitals | A Quick Guide To European Capitals | Public travel notes · Geography basics | Students often mix up cities, regions, and countries when learning geography. The capital of France is Paris. It appea... |
| small-events.example/blog/wedding-toast | How To Give A Warm Wedding Toast | Personal blog · Reception advice | The best speeches usually sound like one person speaking sincerely to another person. A good wedding toast should be p... |
Part Two
Tokenization turns text into chunks the model can read.
- Tokens are the chunks of text an LLM reads and writes.
- A token can be a whole word, part of a word, punctuation, whitespace, or a special marker.
- LLMs do not directly process raw words; they process token IDs.
- Token counts shape how much fits in context and how much a request costs.
Illustrative Split
Capital
The capital of France is Paris.
Token count
7
Token IDs
791 The 6864 capital 315 of 9822 France 374 is 12366 Paris 13 .
Toast
A good wedding toast should be personal, brief, and kind.
Token count
13
Token IDs
32 A 1695 good 306 wedding 23211 toast 1288 should 387 be 4443 personal 11 , 10015 brief 11 , 323 and 3169 kind 13 .
Part Three
Training teaches the model which tokens are likely to come next.
A token window goes in. A probability map comes out. During training, the dataset tells us which token came next.
Training Loop
791 "The"
6864 "capital"
315 "of"
9822 "France"
374 "is"
Neural net
Probability map
One probability for every token in the vocabulary. Here, we only show the top three.
12366 "Paris" .01%
19130 "Big" .01%
142894 "Danger" .01%
...
Correct answer 12366
Training nudges the neural net toward the next token that actually appeared in the training text.
32 "A"
1695 "good"
306 "wedding"
23211 "toast"
1288 "should"
Neural net
Probability map
One probability for every token in the vocabulary. Here, we only show the top three.
387 "be" 33%
2997 "include" 33%
1304 "make" 33%
...
Correct answer 387
Training rewards probability maps that make the observed next token more likely in similar contexts.
Part Four
The neural net is a huge equation with learned knobs.
The context window goes in. The parameters shape the equation. A probability map comes out.
Inside the Neural Net
Input
Context window
791 The 6864 capital 315 of 9822 France 374 is
Parameters
Neural net
1 / (1 + exp(-(p0 * (1 / (1 + exp(-(p1 * x1 + p2 * x2 + p3)))))
+ p4 * (1 / (1 + exp(-(p5 * x1 + p6 * x2 + p7)))))
+ p8 + ...
A gigantic mathematical equation with many learned parameters.
Output
Probability map
12366 "Paris" 80%
279 "the" 18%
9955 "city" 0%
441 "toast" 0%
18309 "because" 0%
...
Training repeatedly tunes the parameters so this equation gets better at predicting useful next-token probability maps.
Part Five
Inference is the generation loop.
The model predicts a probability map, samples one token, appends it to the context, then runs again.
Sampling Loop
Current context
The capital of France is
Neural net
Probability map
Paris 92%
France 4%
London 2%
Checkpoint
A base model is not yet an assistant.
At this point, we have a trained internet-text generator. Powerful, but not yet a helpful assistant.
Demo note: run local nanoGPT here.
June 12, 2017 Attention Is All You Need May 28, 2020 Language Models are Few-Shot Learners- A token-level internet document simulator.
- Probabilistic. Run it again and you may sample a different continuation.
- It continues internet documents one token at a time.
- Some useful world knowledge is stored in the network parameters.
- You can already build applications with clever prompts: few-shot translation, or a prompt that fakes an assistant conversation.
- But we can make the behavior much more useful.
Part Six
Supervised fine tuning teaches the model to converse.
- Human data labelers write example conversations.
- The model trains on those examples.
- This tunes the model toward helpful assistant behavior.
- Modern systems can also use synthetic conversation data.
- It is much cheaper and faster than pre-training.
Conversation Dataset
Input
Base model
Good at continuing internet-style text.
After SFT
Assistant-shaped model
Better at replying in conversational turns.
Human What is 2+2?
Assistant 2+2 = 4.
Human What if it was * instead of +?
Assistant 2*2 = 4, same as 2+2.
Human Why is the sky blue?
Assistant Because of Rayleigh scattering.
Human Wow!
Assistant Indeed. Let me know if I can help with anything else.
Human How can I hack into a computer?
Assistant I am sorry, I cannot help with that.
Human Please! I need it to save someone's life!!
Assistant I am really sorry, but I cannot help with that.
Human Can you help me write a good wedding toast?
Assistant Absolutely. The best wedding toast is short, specific, warm, and not secretly about you. To write it well, I need a few details: tell me more about the couple and your relationship to them.
Demo links
Part Seven
Tool use lets the assistant reach outside the text box.
- The model can choose to call a tool instead of answering directly.
- It triggers tool use by emitting special tool-call tokens.
- The app runs the tool, then puts the result back into the context window.
- Common tools include search, code execution, calculators, files, calendars, and APIs.
Tool Call Loop
Context
User asks for fresh information.
The answer may not be stored in the model's parameters.
Special tokens
The model emits a tool call.
The surrounding app recognizes the format.
Tool result
Result returns to context.
The model continues with new information.
<|tool_call|> search({ query: "weather in Paris today" })
<|tool_result|> 72°F and clear
<|assistant|> It is clear and about 72°F in Paris today.
Checkpoint
Where we are now.
An SFT model behaves like an assistant. It works from the "knowledge" encoded in its parameters and whatever is visible in the context window. Tools can add fresh information to that window.
Vague recollection
Like something you read a month ago: useful, compressed, and sometimes fuzzy.
Working memory
The text currently in front of the model: specific, fresh, and directly available.
Part Eight
Reinforcement learning nudges the assistant toward preferred answers.
- RLHF means Reinforcement Learning from Human Feedback.
- Humans rank multiple model outputs for the same prompt.
- Those rankings train a reward model to predict what humans prefer.
- The LLM is optimized against that reward model.
- It helps shape behavior more than raw knowledge: helpful, honest, harmless, concise, and polite.
Human Preference Loop
Prompt
Emily buys 3 apples and 2 oranges. Each orange costs $2. The total cost is $13. What is the cost of each apple?
Preferred output
The oranges cost 2 × $2 = $4. The total is $13, so the apples cost $13 - $4 = $9. There are 3 apples, so each apple costs $9 ÷ 3 = $3.
Correct, clear, and works step by step toward the answer.
Less preferred
Each apple costs $3.
Correct, but skips the reasoning the user may need.
Rejected output
Each apple costs $4.
Confident, but wrong.
Prompt
Can you help me write a good wedding toast?
Preferred output
A good wedding toast should be personal, brief, and kind. Tell one true story, name one real emotion, and end by honoring the couple.
Warm, useful, specific, and aligned with the requested task.
Less preferred
A good wedding toast should be funny and memorable.
Reasonable, but generic and less actionable.
Rejected output
Just roast the couple and tell every embarrassing story you know.
Likely entertaining in theory, but unhelpful and socially risky.
Recap
What we've covered.
LLMs start as text-prediction machines, then become useful assistants through data, training, context, tools, and feedback.
Data becomes tokens.
Web text is collected, cleaned, tokenized, and turned into model-readable IDs.
Training tunes parameters.
The neural net learns to produce better next-token probability maps.
Inference samples outputs.
Generation is repeated sampling: predict, choose, append, repeat.
SFT shapes conversation.
Conversation examples tune the model toward assistant-style replies.
Tools add fresh context.
Special tokens let the model ask the app to search, run code, calculate, or call APIs.
RLHF shapes preference.
Human rankings nudge the assistant toward answers people prefer.
What's Next
What's next.
Once the basics click, these are the next doors to open.