Re:People paid to train AI are outsourcing their work to AI
by haney on Thursday June 22, 2023 @06:54PM
(#36440333)
The company that I work at does exactly the service that you're describing. We recently spun up a team of Math PhDs to help with data labeling. (https://www.invisible.co/). We're seeing more and more of our clients ask for graduate level data labelers and content creators.
Re:People paid to train AI are outsourcing their work to AI
by dontupvoteme on Thursday June 22, 2023 @11:16PM
(#36442375)
Right now I'm pretty sure just having gpt rewrite the average low to mediocre content that made up the gruel of its generic internet diet and doing fine tunes will get us another 5-10x along, but most hopefully for us little guys out there with a ~24-48GB VRAM cap
If GPT4 really is 8 230M models, the next bit for us will be a few ~1-5M models that swap in for whatever you want to create, or talk about, or what have you
Imagine a model trained just on English football for the purpose of having a good time in the pub that is used when the topic changes to it. I bet you could pass on the dailymails sports page if you add some "u"s into your words.
Or a model finetuned specifically on the library you're trying to debug, maybe even specifically in combination with other tools you're trying to put together.
Can someone in this space invest in doing the hard work to have experts manually curate data?
You know back before Wikipedia, publishers used to pay people to write and edit encyclopedias?
It doesn’t scale. Sure. That’s what the AI you’re building is for though - it will scale.
Throwing compute at ‘the entirety of the internet’ feels like such a lazy way to get what we’re after here.