/
lectern.tv
Play15s30s1x0:00/--:--
DoneWatched
Expand
lectern.tv
33:36/59:48
13:00
13:20
8:50
10:30
14:08
Introduction to Large Language Models
A remarkably dense, hype-free 60-minute mental model of what an LLM actually is—two files on disk shaped by pretraining compression and post-training alignment.
Lecture guide · Written by AI from the lecture. For the lecture’s own words, open the Transcript tab.
Provides a technical overview of large language models, explaining their structure as parameter and code files, pre-training compression, and fine-tuning via Q&A documents. It covers capabilities like tool use and multimodality, future developments in System 2 thinking and self-improvement, and LLM security challenges including jailbreaks, prompt injection, and data poisoning.
This lecture assumes
- Neural Networkstaught in Artificial Intelligence at 43:38
- Deep Learningtaught in Artificial Intelligence at 42:57
- Language Modeltaught in Let's Build GPT: From Scratch, in Code, Spelled Out at 1:36
- Large Language Modelstaught in Artificial Intelligence at 48:50
- Reinforcement learningtaught in Artificial Intelligence at 34:36
- Adversarial Examplestaught in What Can the Theory of Cryptography Tell Us About AI Verification? at 10:08
Key Takeaways
- An LLM consists of a parameter file containing neural network weights and a run file executing the architecture.
- Pre-training compresses massive internet text into parameters via next-word prediction, while fine-tuning aligns the model into a helpful assistant.
- Scaling laws show that next-word prediction accuracy is a smooth, predictable function of parameter count and dataset size.
- Modern LLMs act as operating system kernels that coordinate tools like web browsers, Python interpreters, and image generation.
Concepts Explained22
Large Language Model (LLM)—A large language model is just two files: a parameter file containing the weights and a run file with code to execute them.
Model Inference—Running the already-trained neural network on local hardware to generate text.
Model Training—A computationally intensive process of compressing a large chunk of internet text into network parameters.
Next Word Prediction—A neural network task where given a sequence of words, the model predicts the most probable next word.
Examples, Proofs & Derivations
Works & Sources Cited10
Paper
InstructGPT—OpenAI paper detailing human labeling instructions and alignment.
Paper
Sparks of Artificial General Intelligence: Early experiments with GPT-4—Bubeck et al., 2023, cited for evaluation benchmarks across knowledge domains.
Primary Document
Her—Movie cited as an analogy for speech-to-speech AI interfaces.
Book
Thinking, Fast and Slow—Book by Daniel Kahneman, populating the System 1 vs System 2 cognitive framework.
Paper
Tree of Thoughts: Deliberate Problem Solving with Large Language Models—Paper by Yao et al., 2023, cited for Tree of Thoughts.