• PyCon US took place in May and 154 videos started being published on YouTube. Many other interesting conferences also happened during this time.
• People's time is limited, so they can't watch all the videos. One option is to guess the interesting content by looking at titles and descriptions, but this often leads to disappointment.
• There is a need to find the best videos to watch based on "needs", extract the main teachings of all content, and store it in a consumable/searchable way.
• In 2024, digital assistants can help with this. The author used "ollama" to run AI models locally (relying on llama3:8b due to limited machine resources) and "fabric" to provide prompts and download transcripts from YouTube videos.
• An example output is shown for the video "Iterate, Iterate, Iterate! (Moshe Zadka) — PyTexas 2024", including summary, ideas, insights, quotes, habits, facts, references, one-sentence takeaway, and recommendations.
• After evaluating the results, it is clear that the extracted content is not 100% accurate, with issues like focusing on a small part, highlighting superfluous stuff, and misinterpreting. Possible reasons are the chosen model and relying on transcripts (missing visual information).
• The author wrote a quick script to run information extraction on YouTube playlists and created a repository (https://github.com/dethos/con... to store the results. People can ask for specific conferences to be added by creating a new issue.