from Hacker News

Show HN: Askyoutube – Ask YouTube anything

by machinelearning on 6/16/23, 8:29 AM with 15 comments

Hi everyone, I built this recently and found myself and my friends using it for various different things so I figured I'd share it here.

It uses LLMs to answer queries from videos so you don't have to watch them yourself! Each video found is also included with a summary of its transcript for further exploration. Most importantly, it uses up-to-date videos and can answer question about real and recent things unlike ChatGPT.

It is particularly useful for:

1. Getting up to date information about the news from different sources E.g.

- What are the features of the recently released Apple Vision Pro?

- How many people were evacuated from the recent eruption in Hawaii?

2. Product reviews E.g.

- What's the best vegan & cruelty-free skin cream?

- What is best GPU for deep learning in 2023?

3. Learning about a specific topic E.g.

- What are the requirements for the commercial pilot's license?

- How do I do a cartwheel?

Let me know what you tried and submit feedback in the interface, I read all of them! Also do include your email if you want a follow up for your feedback

Email: askutubeai@gmail.com

  • by WA on 6/16/23, 11:38 AM

    "request might take up to 5 minutes to process" -> in a world of broken AJAX calls and hence, broken spinners, 5 minutes is literally forever to know if this thing works at all. If I don't see a result within 5 seconds, I assume that my internet connection hangs or the server just broke a thing.

    You probably have to give some other indicator to show the user whether your site is working at all or not.

  • by HanClinto on 6/16/23, 12:01 PM

    Wow, I just tried this out with a query that is answered by a video that I recently made. It showed not only my own video, but also several others I hadn't yet seen.

    Previously, I had already tried using YouTube's searching tool to find videos for this, but didn't come up with much. Your tool returned several results I had not seen before, and they look super laser-targeted. REALLY impressive work! This is insanely cool, and FAR better than YouTube's search algorithm. Very very well done!

    I'm interested in doing a similar YouTube indexing project (but indexing with pose detection rather than semantic segmentation) -- do you have any guidance for people wanting to do a similar project?

  • by ggordbegli on 6/16/23, 12:26 PM

    This is wonderful. I made questionyoutube.com yesterday to solve a similar problem, I think this is better in the limit though.

    If you can just get search time down, and make the summarization a little better, that would be amazing.

    For instance, I tried this with 'what Toyotas should I avoid' and it found good relevant videos, but in the summary of 'Bad Toyotas you need to avoid' it misses that certain Toyotas from 2007-2009 had an oil burning issue, which was the main topic of the video. Querying the transcript of the video with 'what Toyotas do I need to avoid' gives a more complete answer. I'm curious how your handling the initial summary.

  • by mdrzn on 6/16/23, 12:42 PM

    "Note: request might take up to 5 minutes to process."

    I'd say show me some progress?

    Write down some text like "searching for videos.. summarizing.. finding answers.." so we know when it's moving forward.

  • by HanClinto on 6/16/23, 11:46 AM

    Really impressive!

    I'm guessing you're using LLMs to generate vectors for the search query, and then using that to index into your database of vectors generated from the transcript of each video? Is that correct?

    If so, which LLM are you using to generate the vectors? What does your backend look like? How many videos are currently in your database?

  • by freediver on 6/16/23, 8:48 AM

    Glad to see this resurface as an idea. We experimented with this in 2019, you can check the video of the demo here:

    https://vimeo.com/754908681

    Hope it inspires some features.

  • by aerio on 6/16/23, 12:14 PM

    230 days ago, this same idea got zero traction on HN lol

    https://news.ycombinator.com/item?id=33370718

  • by Graffur on 6/16/23, 10:46 PM

    5 minute wait lol.. at least show some pre-loaded examples