This document is a transcript of administrative and technical discussions from the '#admin' channel of the "Solve It With Code" Discord server. The conversations, primarily featuring course creator Jeremy Howard alongside teaching assistants and students, cover a wide range of topics related to the course logistics, the "SolveIt" learning platform, and community management.
Key logistical areas addressed include troubleshooting website access, clarifying course schedules and costs, and managing student requests. A significant portion of the discussion is dedicated to the technical architecture of the "SolveIt" web application, revealing its foundation on FastHTML and a per-user SQLite database system, hosted on a VPS and backed up via cron and rsync.
Community management is another central theme, with extensive dialogue about coordinating live session times across global timezones, leading to the organization of community-led viewing parties and the creation of timezone-specific channels. The transcript also captures real-time user support for the SolveIt platform, addressing bugs, performance bottlenecks, and feature requests. A notable sub-thread details a user's attempt to use Google's Gemini models to summarize the course videos, highlighting practical challenges with large context windows and the general inadequacy of current AI tooling for such tasks.
Key Administrative and Technical Discussions from the 'Solve It With Code' Discord
Nov 9: Initial Platform Issues & Course Access: Users report and resolve initial website access and sign-up issues. Jeremy Howard clarifies that the course platform will be accessible via the sign-up website after November 24th.
Nov 10: "SolveIt" App Tech Stack Revealed: In response to user curiosity, Jeremy Howard explains the course's web app is built with fasthtml and fastlite. The architecture uses a single SQLite database for user sign-ups and a separate SQLite database for each user's course work. The platform is hosted on a bare metal or VPS server, with backups managed via a cron and rsync script.
Nov 19: Live Session Timing & Global Audience: A discussion begins about the live session schedule, which is challenging for attendees in different timezones (e.g., 4:30 AM in India, 1:00 AM in parts of Europe). To accommodate this, fast.ai teaching assistant Audrey M. Roy Greenfeld proposes hosting "viewing parties" for the class recordings.
Nov 20: Organizing Viewing Parties: The community discusses logistics for the viewing parties, including alternative timings and how to host them using Discord's "Watch Together" feature. An official event is created for the first session.
Nov 21: Timezone-Specific Channels Created: To help organize regional study groups and viewing parties, Jeremy Howard creates dedicated timezone-based channels for different regions, including Australia and Africa.
Nov 25: Platform Scope and Data Privacy: Jeremy Howard clarifies that the SolveIt platform currently uses the Anthropic Sonnet model and is intended for learning and hobby projects only. Users are advised not to share private or commercial data.
Nov 27: Platform Updates and Bug Fixes: An update to the SolveIt platform is announced to support both Ctrl and Cmd as shortcut keys. To get the update, users must use the "stop solveit" link on the homepage and then connect again to get a new instance.
Nov 28 - Dec 9: Troubleshooting and Feature Requests: Users report various issues, including the platform freezing, hanging on computation, or showing blank screens on expired URLs. Common causes are identified as long-running computations or infinite loops. Users request a "stop thinking" button and a centralized timetable.
Dec 11: Accessing Course Materials: It is clarified that links to past lessons are available on the SolveIt dashboard. Code and gists from lessons are shared in the relevant daily channels (e.g., #day-2).
Dec 11: Summarizing Lessons with AI: A detailed discussion unfolds about using AI models like Gemini to summarize the unlisted YouTube lesson videos. The effort highlights significant practical challenges, including the massive token count of video/audio files (exceeding 2M token context windows), the difficulty of automating video splitting, and the generally poor state of AI tooling for such real-world tasks.