Hey guys! Over the past weeks, I've been working like crazy on my newest product: Ortus (it's currently a Chrome extension for RU-vid). You can install it here: chrome.google.com/webstore/detail/ortus-your-youtube-ai-bud/jmpepfdhkjkknfpnfohnmnjoceepcbmp It basically saves you time and helps you learn faster, by allowing you to ask your AI buddy questions about the videos while enjoying your favorite RU-vid channels (including yours truly!) I believe that in not so distant future we'll have much better ways of consuming information. Here are some questions you can currently ask it: * "Which books did Lex recommend in this video" and get an immediate reply and a precise timestamp of that moment. * "What are the key insights from this video?" * "What are some similar videos you could recommend me?" * "What happened at the 23rd minute of the video?" * Or just ask "What can you do?"! :) and much much more.
Hi Aleksa! I was recently thinking about this kind of chrome extension for educational purposes and here you are with this amazing product. I had a lot of ideas for my own project and after trying out many different things on Ortus I’d like to share some improvements. 1. Currently, it’s not possible to see the exact (AI-generated) transcript of the videos on Ortus chat. This should be a fundamental feature as the summaries may skip some important details and the youtube api already offers that, you just need to integrate the transcript. In addition, the use of youtube transcripts should enable you to add minimal support immediately, at least for the shorter videos on youtube. 2. Ortus cannot identify you (addresses as “the speaker”) in the chat, even though you mentioned your name (though the summary addresses you as “Alex”). But, it can do so if names are clearly mentioned (works in some other videos). Maybe this can be solved if more data (like channel name, video title, description data, and general knowledge from all other videos of the channel) are considered. Would be nice to see a dialogue-like summary if there is more than one speaker in the video. 3. It’s important to give access to the Internet. For instance, the LLM you’re using has no info about “Aleksa Gordić”, and you may solve this by utilizing the social media links provided in the “about” section. Since you’re making this product for educational purposes, google searches, Wikipedia links, etc will be really helpful. 4. Sometimes it is difficult to get the time stamp of a section or the probable time segment may not be accurate for the section. Perhaps, further RLHF training can help improve this easily. Moreover, the context window needs to incorporate data from the last three chats at least, otherwise very detailed questions are needed to be asked every time, and cannot just follow along with the present context. 5. It would be great if you add coding support (information should be extracted from the github/colab notebook link provided), otherwise the general question answering doesn’t help much in the coding videos. You can add the option to open the chat in a new window just for more space to see the code-based answers clearly. This one is a bit too ambitious idea, but, hopefully, you can make it possible in the long run. I don’t know what kind of a business model you currently have in your mind, but I feel it’s best if you reach out to youtube channels and offer this service through contracts. This product would create a lot of value for many educational channels and surely contracts would lead to a much more sustainable income than earning from user subscriptions. Providing support to random videos just from the preferences of the current users will never allow Ortus to get sufficient recognition. On the other hand, you can actually use the contracted youtube channels to promote Ortus and this will exponentially increase the number of users. Moreover, you can still keep the option for a user subscription to give access to some premium features (like conversation history on a video from previous views, some statistics about learning progress, etc). Lastly, I’d like to stress that you must expand support to smartphones as soon as possible because a Chrome extension alone cannot help you reach the vast number of users who really need this product. My personal idea was to have an app on smartphones that would have a floating chat pop-up over the youtube app. In this way, you can undoubtedly convince Y Combinator for the funding. Best of Luck! P.S. Discord is fine for creating a community fast. Still, I think you should also have a separate GitHub repo just for the bug fix-related issues (found a bug: it’s not possible to undo a mistakenly upvoted video) and for getting suggestions for upgrades.
@@xl0xl0xl0 can't be done tried to do it :/ I guess it makes sense from the safety side: RU-vid doesn't allow its creators to edit the videos in case they replace the audio with something malicious, etc. Otherwise I'd immediately fix it. I hate it...
Great content Aleksa! This is a really cool project, and definitely amazing! Side Note: I found it a bit difficult to watch the video from headphones, as your video is hearing only from left side (at least in my case).
Cool product! Are you using Whisper to get the transcripts and then GPT-4 API for the question answering based on the transcripts, or did you fine-tune/train from scratch any models on your own?
Would be nice to have a LLM trained on ALL of the content of a given youtube channel, and chat with it and also to be able to ask about where (which video/timestamp) is the information that I am asking.
Great project! Since you are using ChatGPT probably. You can basically offer translation for non English speakers, where they ask their questions in their mother tongue to better understand topics with complex vocabulary. This is indeed a very helpful tool. I am curious why it works only on these 4 channels. Are you preprocessing the videos in the backend to generate summaries for each video? Will this work with very recent uploads ( less than an hour)? Why not just feed the network the captions of a video and make it work for basically any channel with scripts (and you can test on auto generated captions as well)?
I am currently using "RU-vid and Article Summary powered by ChatGPT" plugin, which served me iwell so far. This UI looks great and I look forward to trying it out. One question: Do you save the selected transcripts in vector DB? There is no way to fit one in real time with current token limit.
they mostly focus on summaries, that's not our primary focus but you can ask the agent and it'll give you summaries as well :) (plus we do have the summary window below which is about to become 10x better in 10 hours :) )
@@TheAIEpiphany Please also add the option to exchange the positions of the chat and the summary window because it is difficult to follow the video when I'm reading the summary, for which I need to scroll down
Well done. I wonder about the legal implications. I assume not every RU-vid content creator would love his videos to be accessible in this manner as it can significantly reduce the view time. Do you need to get permission form a RU-vid channel to include it into this AI tool.
Thanks for the feedback, we'd love to see how we can support your use-case, join the Discord and describe it we'll see what we can do about it! Link is here: discord.gg/fNP5apexWw
I definitely understand that there’s a lot more to it. But looking at the product it looks like its a chat fine-tuned model. So I was just curious to know whether it was oasst, chatgpt, or some open-source instruction fine-tuned model (RedPajama, MPT, …)