Proposal: Youtube Video content assistant

Youtube Video content assistant
Proposal in one sentence: Web app that takes a youtube link, retreives the transcript from it, cleans up the text and uses GPT-3 to answer questions, write discussion posts, etc

Description of the project and what problem it is solving: When I am watching an informative youtube video - either for class, work or a project - I often copy and paste the transcript from the video into OpenAI’s GPT-3 playground to help me answer quesitons about the video and generate deliverables like discussion posts and short essays. By automatically retreiving the transcript and pre-configuring the optimal settings, it could be a very useful tool
Grant Deliverables:

Web app that answers questions about the content of any youtube video, and generates variable length blog/essay posts related to it

Squad Lead: Curran Janssens
Discord: cjanssens#2932


Hey, the idea sounds super interesting. Why specifically would you need GPT3 for this? Won’t a regular BERTQA model solve the existing issue of finding answers? Or maybe if we go a bit primitive, even nltk has robust functionality to do this

1 Like
1 Like

Thanks for the feedback! Smaller models like you suggested would definitley make sense (and be cheaper!) for simple tasks like retrieving content or answering simple factual questions, but GPT-3 can create much longer form content, and is much higher quality for more complicated topics where some of it’s latent knowledge is very useful as well. I’ve found GPT-3 replacing a lot of my writing in a way I wouldn’t expect smaller model to be capable of

This sounds like a good plan and a very useful system for a lot of people! Some clarifying questions that I hope are helpful in scoping out the project:

Do you mean you will employ OpenAI’s GPT-3 API to build the webapp or is there a possibility you could use open-source alternatives like Eleuther AI’s models that they’ve publicly released? I have no prior experience deploying them at scale though, so I wonder if that’s prohibitively expensive and if so then the OpenAI API approach certainly looks more appealing for a first iteration.

In that case, have you considered how many API calls you get for free and at what point you’d have to start paying for such a service beyond the grant support?

Lastly, could a browser plugin be better suited (and faster to develop) to this task, since it is meant to generate output from video transcripts?

Great Feedback!

I plan on using OpenAI’s GPT-3 API mostly for performance and convenience reasons - it’s easy to set up and configure, and for a broad task like this a higher quality model is worth it. I don’t think costs will be a huge issue, and if they are I would probably just switch to a smaller GPT-3 variant.

I hadn’t thought about a brower plugin but that is a really good idea. I don’t have any experience making extensions but if I find someone who does I might pursue that, thank you!

1 Like