Alvin Wan is a Sr Research Scientist at OpenAI. Previously he was a Senior Research Scientist at Apple working on large language models with Apple Intelligence. He formerly worked on Tesla AutoPilot and graduated with his PhD at UC Berkeley with 3000+ citations and 800+stars for his work.

Zao Yang is currently a co-founder of \newline. Previously he co-created Farmville (200m users, 3B in revenue) and Kaspa ($3B market cap). He’s self taught in generative AI, deep learning, and ML. Newline have deployed internally multiple generative AI apps for sourcing, tutoring, data extractions. \newline has over 250k professionals that include Salesforce, Adobe, Disney, and Amazon go to level up when they need to learn quickly and efficiently.

Dipen Bhuva. 5th year PhD Grad Student with 100+ citations and 16 published research papers. PhD Researcher at Cleveland State University. Northeastern University graduate with Masters in Informatics.

Philosophy in building the course:

• Start with hello world with a model
• We start with evergreen concepts in transformer based large language model concepts (tokenization, state of art tokenizers, semantic understanding of embeddings)
• What you build is applied to a state of the art model Llama
• Certain concepts we dive deep into make sure they stick
• Introduce you to state of the art concepts but won’t go deep. There will be code and techniques or a recipe book of different techniques there.

Unit 1: Introduction

• Week 1: Build an LLM Inference API

  • Overview of Terminology
    • Note: ChatGPT vs. LLM, two phases of inference (prompt processing vs. autoregressive decoding), three phases of training (pre-training, post-training instruction-tuning, post-training fine-tuning), compression (quantization, pruning), RAG
  • Types of AI applications
  • Categories of LLM Knowledge
  • Phases of LLM
  • RLHF
    • How RLHF Contributed to ChatGPT's Success
    • RLHF Phase
  • LLMs Powering Everyday AI Products
    • How Chatgpt differs from other LLMs
    • Training Data Comparison
    • Model to use daily
    • DeepSeek Distill Models
    • DeepSeek Chain of thought
    • TestTime Compute
  • Inference: Prompt Processing vs. Autoregressive Decoding
    • Prompt Processing
    • Autoregressive Decoding
  • RAG: Bridging LLMs and Information Retrieval
  • Ecosystem
    • Note: Vector DBs to support RAG, inference APIs for pre-trained models, libraries and datasets for fine-tuning. System is setup for you to consume and tune pre-trained models.
  • Lifecycle of a Model
    • Note: Overview of course effectively. We will build everything from scratch.
  • “Hello World” LLM inference with Huggingface
    • Note: Small model like opt-125m on Google colab.
  • Build a “Hello World” LLM API
    • Note: Use Modal for deployment. This makes it simple. No need to make a Flask app, as server-less deployment means you just need to define a function.

• Unit 1.1: Introduction to AI applications

  • Introduction to AI applications
    • LLM writing
    • Voice
    • Images
    • Video
    • Multi modal
  • Intro to AI modalities
    • Q&A
    • Image generation
    • Briefings and reports
  • Examples of AI applications
    • Industry AI applications
    • Role based AI applications
  • Technologies
    • RAG
    • Prompt engineering
  • How to brainstorm AI applications
    • What are you doing at work?
    • What are existing SaaS workflows
  • Vertical applications and their example technologies
  • Indie hacker examples
  • Venture funded LLM examples

Unit 2: Building a Shakespearean language model

• Week 2: Understanding tokenization
  • What is a tokenizer? What makes a tokenizer “good”?
  • Build a baseline ‘word-based’ tokenizer
  • Build a byte-pair encoding
  • Compare with state-of-the-art tokenizers (e.g., Llama’s)
  • How can Byte Latent Transformer put an end to Tokenization
• Week 3: Understanding Token Embedding
  • What is an embedding? What makes an embedding “good”?
  • Understand the semantic meaning of word embeddings
  • Train a word embedding on Shakespearean tokens
  • Note: Train as word2vec does, iteratively.