from Hacker News

[Help] Which AI Model to Use?

by deepshaswat on 8/26/24, 12:33 AM with 4 comments

Anyone here has done a custom ML setup or used an AI model deployed in local or AWS

The requirement is I have 700k records of YouTube channels with descriptions and keywords. I also have a list of categories and subcategories with description: 776.

Now I want to map each account to a relevant category and sub-category and not spend more than $50-100.

I thought to use AWS Bedrock with Titan model but the cost comes around $11k.

Any pointers will be really helpful

  • by any_throw777 on 8/26/24, 1:56 AM

    I hate to say it depends, but, it depends.

    If you have local hardware (16-24 gb of video card), you should be able to use a number of different ml approaches, including LLMs. You could also try a forest classifier, or other approaches, but without specific about your data, it might be a little challenging to rec one.

    I would try OpenAI or Anthropic. You should be able to fit all your categories in a prompt, and then ask the specific item. Something like >>

    You are a classifer agent. Here is a list of categories and subcategory.

    -item 1 -item 2 -... -item 776

    The details of the channel are x, and here are a few descriptions.

    --- The above could work if you can fit all the categories and description in one prompt. Both OpenAI and claude are pretty large.

    The downside of this approach is that it's very hard to predict the classification with any reliability.

    The upside is that it's fast and very likely to fit your budget. GL

  • by wizzerking on 8/26/24, 2:19 AM

    what you want is a video summarization net possibly followed by an LLM PaperswithCode or arxiv to the resue just search for pages with open source code or got directly to github. Like deepshaswat says inference requires no specialized hardware, but training can be a problem unless you have a PCIE 4 mainboard with an RTX 4090 with 20 GB of RAM I bought this system with core i9 and 64 Gig Mainboard RAM for 5K 3 years ago, you may be able to only pay 2-3K nowadays. Better performance with Radeon and Threadripper is probable from the stats I see on Tom's and overclockers Good Luck