Intro to Large Language Models (LLMs)

Trace the shift from hard-coded rules to deep learning and transformer-based large language models, enabling generative conversations about flowers and Iris data set–based image classification.

Explore why recurrent neural networks struggle with long text and context memory. Discover how transformers enable parallelization and attention to improve language understanding.

Explore how the attention mechanism and transformers weigh input tokens to compute a weighted sum, enabling long-range dependencies. See self-attention in action through a translation example of French to English.

Explore the transformer architecture, including encoder-decoder design, and how transformers process all input words at once, contrasting with RNNs and detailing encoder blocks.

Create input embeddings by tokenizing text, mapping tokens to vocabulary indices, and retrieving word embeddings, then apply positional encoding and optional padding or truncation for the transformer encoder.

Explore how the encoder block uses multi-headed attention to compute token context via query, key, and value vectors, dot-product similarity, scaling, softmax, and weighted sums.

Explore how the feed-forward layer in the transformer encoder processes per-token representations after self-attention, applying linear projections, an activation, and a second linear transformation to capture non-linear relationships per token.

The lecture explains how masked multi-head attention in the decoder uses encoder embeddings and prior outputs to predict the next word, hiding future words to train sequence generation.

Explain how transformer decoders use encoder outputs to predict the next token, forming a context vector and probability distribution through feed forward and softmax.

Explore what GPT stands for and how ChatGPT uses a generative pre-trained transformer to generate human-like text, powered by pre-training and the transformer architecture.

Trace the evolution of large language models from GPT-1 to GPT-4, highlighting scaling, performance gains, and ChatGPT's conversational tuning and API access.

Learn to generate text with a GPT model by building a simple function that calls the OpenAI completion API, passing a prompt and parameters like max tokens and temperature.

Experiment with max tokens of five, twenty, and fifty and with temperature 0 or 1 by turning them into function arguments to see how prompts shape GPT output.

Learn to build a text summarizer that extracts keywords from long text blocks by prompting a language model with system, user, and assistant messages.

Code a simple chatbot that responds poetically by using a system message, example prompts, and the OpenAI chat completion API with GPT-3.5.

Explore how Lang chain lets you import your own data and enable language models to reference it when generating responses, using the Lang chain framework in Python.

Explore how Hugging Face's Transformers library enables loading, tokenizing, and generating text with pre-trained large language models, and fine-tune them for sentiment analysis, named entity recognition, or text classification.