As far as I can tell, this is what happened.

Abstract:

This video documents a lesson (Lesson 5) on software development using the Solvit platform, covering three main practical examples. First, Jeremy Howard demonstrates a full development cycle by implementing a feature request for the cozette library: adding support for new models via a third-party provider (Groq). This segment covers cloning, interactive development within Solvit, and releasing the new version using nbdev. The second part, led by Jonno, continues building a "Solvit-like" web application. He implements logic to differentiate between Python code (for execution) and natural language prompts (for an LLM), and then adds multi-user state management using browser sessions. The final segment has Howard introducing database persistence using SQLite and the fastlite library to store user dialogues. He concludes by creating a Pythonic wrapper library for the Daisy UI framework, automating the component generation process with an LLM.

A Practical Guide to Iterative Development with Solvit and nbdev

• 0:00:49 Student Work Showcase: The lesson begins by featuring two student projects: one using Solvit for an in-depth exploration of the fastai source code, and another creating a "starter" dialogue for learning about biology.
• 0:05:17 Live Feature Implementation: Jeremy tackles a real GitHub issue for the cozette library—adding support for new GPT-OSS models. He decides to implement this using the Groq API due to its high inference speed.
• 0:13:36 Modifying Library for New Providers: He modifies the library's Client class to accept a base_url and an environment variable name for the API key. This change makes the library flexible enough to support any OpenAI-compatible API, not just Groq.
• 0:25:27 End-to-End Release with nbdev: After verifying the changes, Jeremy demonstrates the full nbdev workflow: cleaning the notebook (nbdev_clean), exporting it to a Python library (nbdev_export), committing the code, creating a GitHub release (nbdev_release_gh), and finally publishing the new version to PyPI.
• 0:35:39 Building a "Solverish" App: Jonno continues building a chat application inside Solvit. He introduces logic to automatically distinguish between user inputs that are valid Python code and those that are natural language prompts.
• 0:38:26 Dynamic Execution Path: The application is modified to execute valid Python in a shell and send natural language text to an LLM for a response, creating a hybrid code/chat interface.
• 0:49:39 Multi-User Support via Sessions: To handle multiple simultaneous users, Jonno refactors the app to use browser sessions. This stores a unique user ID in a cookie, allowing the backend to maintain separate chat histories and execution environments for each user.
• 1:23:51 Adding Database Persistence: Jeremy takes over to replace the ephemeral, in-memory user data storage with a persistent database. He uses SQLite via the fastlite wrapper library.
• 1:25:50 Iterative Database Schema Design: He demonstrates creating tables for users and dialogs, emphasizing that the schema can be easily modified as new requirements emerge by using transform=True during table creation. He stores the ordered list of messages as a JSON string within the dialogs table.
• 1:58:11 Creating a UI Component Library: Unhappy with the verbosity of using Daisy UI's CSS classes directly, Jeremy creates a Pythonic wrapper library (fast-daisy). The goal is to provide a cleaner, more consistent API for building UIs.
• 2:01:40 Automating Library Creation with an LLM: In a clear demonstration of practical AI use, Jeremy scrapes the Daisy UI documentation website, uses an LLM to parse the HTML into a structured list of components, and then instructs the LLM to write the Python code that generates the entire wrapper library automatically.