00 · What Is a Large Language Model?

GENE 46100 — Unit 01

Haky Im based on 3Blue1Brown LLM Explained

2026-04-11

One sentence definition

An LLM is a function that assigns probabilities to the next word in a sequence — not one answer, a whole distribution.

Source: 3Blue1Brown — Large Language Models Explained Briefly

Part 1: How a Chatbot Works

Step 1: The setup — a system prompt

Before your question, the model receives a description of its role.

Source: 3Blue1Brown — Large Language Models Explained Briefly

Step 2: The incomplete script

Your question is appended. The AI’s response is missing — the model must fill it in.

Source: 3Blue1Brown — Large Language Models Explained Briefly

Step 3: Predict one word

The whole text so far is fed into the model. It predicts the most likely next word — here, “used”.

Source: 3Blue1Brown — Large Language Models Explained Briefly

Step 4: Append and repeat

“used” is appended to the script. Feed it all back in. The model now predicts “to”. Repeat.

Source: 3Blue1Brown — Large Language Models Explained Briefly

The full chatbot loop

The model is just completing the script — word by word — until the assistant’s turn is done.

Source: 3Blue1Brown — Large Language Models Explained Briefly

Same prompt → different response

The model doesn’t always pick the most likely word. Intentional randomness makes outputs feel natural — and means identical prompts yield different responses.

Part 2: Training

What is the model trained on?

Vast amounts of internet text. Reading GPT-3’s training data would take over 2,600 years of non-stop reading.

Source: 3Blue1Brown — Large Language Models Explained Briefly

What is a “parameter”?

The model is a function with hundreds of billions of tunable dials — parameters. They start at random values and are adjusted through training.

Source: 3Blue1Brown — Large Language Models Explained Briefly

Billions of parameters

Each dial is a continuous number. Their collective setting determines everything the model knows.

Source: 3Blue1Brown — Large Language Models Explained Briefly

The training objective

Show the model text with the last word hidden. Ask it to predict. Compare against the truth.

Here: model guesses “age” (87%) when the correct answer is “worst” (10%). It’s wrong.

Source: 3Blue1Brown — Large Language Models Explained Briefly

Backpropagation fixes that

Adjust the parameters so “worst” becomes more likely. Do this across trillions of examples.

After training: “worst” jumps to 70%.

Source: 3Blue1Brown — Large Language Models Explained Briefly

Two phases of training

Training is not one step — it has two distinct phases.

Step 1 — Pre-training: predict next words on raw text → learns language Step 2 — RLHF: human feedback shapes it into a useful assistant

Source: 3Blue1Brown — Large Language Models Explained Briefly

RLHF: Reinforcement Learning from Human Feedback

Human workers review model outputs and flag problematic or unhelpful responses. Those corrections adjust the parameters further.

This is what turns a raw language model into a chatbot.

Source: 3Blue1Brown — Large Language Models Explained Briefly

Part 3: The Transformer Architecture

Why GPUs?

Training requires billions of parallel multiply-and-add operations. GPUs — originally designed for graphics — are built for exactly this.

Source: 3Blue1Brown — Large Language Models Explained Briefly

The 2017 breakthrough: Transformers

Before 2017, models processed text one word at a time (sequential). Transformers process the whole sequence at once (parallel).

Source: 3Blue1Brown — Large Language Models Explained Briefly

Words → vectors

Each word is mapped to a long list of numbers — a vector. These numbers encode the word’s meaning in a form the model can compute with.

Source: 3Blue1Brown — Large Language Models Explained Briefly

Attention: vectors that communicate

The attention operation lets every vector update itself based on context from other vectors.

“bank” surrounded by “river” and “jumped into” → the vector shifts to encode riverbank, not financial institution.

Source: 3Blue1Brown — Large Language Models Explained Briefly

Over and over: layers refine meaning

Attention and feed-forward layers alternate, repeatedly. Each pass enriches the vectors.

Right: the “bank” vector travels through meaning-space — from the raw token, to riverbank, to narrative context.

Source: 3Blue1Brown — Large Language Models Explained Briefly

The final prediction

After all layers, the last vector encodes everything: context + training knowledge. A final function converts it to a probability distribution.

*“Down by the river bank … until they jumped into the ___“* → water: 51%, river: 19%, lake: 7% …

Source: 3Blue1Brown — Large Language Models Explained Briefly

Part 4: Why This Matters for Genomics

DNA as a language

The exact same framework applies to genomic sequences:

Language LLM	Genomic LM
Words / tokens	Nucleotides, k-mers
Next-word prediction	Next-nucleotide prediction
Pre-train on internet text	Pre-train on reference genome
Fine-tune for task	Fine-tune for TF binding, promoter activity

What we’ll build in this unit

1. slides-01 — Tokens, embeddings, dot products, softmax — the geometric foundations
2. slides-02 — Transformers in detail: attention, Q/K/V, multi-head, residuals
3. slides-03 — Training a language model: tokenization, loss, the training loop, generation
4. notebook-01 — Build nanoGPT from scratch on Tiny Shakespeare (Karpathy)
5. notebook-02 — DNA language model: pretrain on the human genome, fine-tune for prediction

Key Takeaways

• An LLM outputs a probability distribution over the next word — not one answer
• Chatbots work by autoregressive generation: predict → append → repeat
• Parameters begin random; backprop over trillions of examples shapes them
• Pre-training on text + RLHF from human feedback = a useful assistant
• The Transformer (2017) processes the whole sequence in parallel via attention
• Words become vectors; attention lets those vectors update each other from context
• Behavior emerges from training — nobody programmed what the model knows