Рет қаралды 10
This is behind-the-curtain video about how AI-generated analysis of AI research is made on this channel. I strongly recommend watching at speed x2. The script that is used here:
1. Checks latest articles on arxiv.org, stores them in local database (so that I track which articles were already processed)
2. Describes these articles in simple terms, throwing in some real-life examples and applications
3. Creates intro, outro and all the metadata (like youtube title, video description, tags etc) by a chain of prompts technique
4. Converts text to audio
5. Creates lip-synced avatar based on audio
6. Creates full sized youtube video based on avatar and research articles
7. Uploads the video to youtube (with all the metadata filled in)
There is also an alternative script which selects the most interesting article (based on some specific criteria) and creates in-depth analysis of the article's content (for purpose of research compilation, only the title and abstract are analysed, but in the in-depth analysis the whole PDF is parsed and thought through). I created a loop of question-and-answer about the PDF content, throwing in full PDF into context (RAG with some prompt engineering can be used to reduce amount of tokens, however at this point it was not worth the effort). It produced extremely satisfying responses in my tests (both for long and short research papers).