Night School
Build a Neural Network From Scratch: Training, Fine-Tuning, and RL (Part 2 of 2)
About this lecture
Everyone has access to the same models and the same harnesses. What is increasingly yours alone is your data — and the moment your data is the IP, someone on your team has to understand how a model gets trained on it. This session closes the September series by building a small neural network from scratch, so every piece you have been importing becomes something you have written once yourself. Aaron Gallant walks through what a training pipeline really does with a corpus of data, what hyperparameters actually control, and where fine-tuning ends and reinforcement learning begins. We will work through the notebooks assigned after part one together. The goal is to demystify training and move you from using AI to making it.
This Night School runs in two parts — same teacher, two sessions two weeks apart, with work to do in between. Part one on September 2 covers what is actually happening inside the model; part two on September 16 builds one from scratch. The two are designed to be taken together — bring the notebooks from part one.
Note on prerequisites: this is a crash course, not an introduction to programming. If you are already a little bit technical you will be fine. If you have never written code, this one will be a hard start.
In this lecture
- Build a small neural network from scratch and see every component you have been importing
- Understand what a training pipeline does end to end, from a corpus of data to a trained model
- Learn what hyperparameters are and how those configuration choices change what you get out
- Know the difference between fine-tuning and reinforcement learning, and what a reinforcement learning environment is for
- Walk out with the foundational knowledge to test-drive training a small model of your own
About the instructor
Aaron Gallant — Lead Instructor at Gauntlet AI, with Gauntlet since the beginning. He has spent 18 years in technology across software development, data science and analysis, data pipelines, and internal tooling, including five years at Google building tools for internal support workflows and operations. He works on on-prem models and has run training pipelines himself, including generating synthetic data to train HIPAA-compliant models for clinical research. He holds a master’s degree in Computer Science from the University of Illinois at Urbana-Champaign and a bachelor’s degree in Political Science and Philosophy from the University of Rochester. As Gauntlet AI CTO Ash Tilawat put it: “If there’s one person that truly understands the foundations of classical ML and how they translate into LLMs, that’s Aaron.”