Slop Bucket Idea – a dataset of AI slop (train AI what not to do)

2 pointsby IAmNeo2 mesi fa4 comments

I just had this idea, you read it all the time AI slop is so prevalent people are getting banned for a year for submitting science papers to arXiv with it, moans of angst from developers, even Microsoft doing its own study where AI degrades the quality of simple documents, and the beloved em-dash.

I don't really have the know-how or the time but it occurred to me, if we created a public data set that could be submitted to publicly, we could catalog and organize all the AI slop, the different types, with explanations about why it is slop and why not to do it, and then train a large language model using this data set included, to help correct itself.

I don't really know the technical details of training a large language model,is this even possible?

4 comments

HerbManic2 mesi fa

I suspect this is why places like Reddit are so valuable. With the voting system, they can correlate AI output with peoples approval. They are essentially ranking how well the output is received.

This doesn't mean this stuff will improve drastically as it might just highlight how low the bar is for some folks, but it is better than nothing.

Sort of like how reality TV was considered broadcast anesthetic slop, and yet people couldn't get enough of it and it is still with us decades later.

IAmNeo2 mesi fa

I like this, if there was a way to collate all the tagged slop from reddit, and somehow tell the large language model not to implement them but to actively avoid it's patterns.

Is slop considered a shortcoming of AI not being intelligent/advanced enough yet?

ares6232 mesi fa

Most/some of what we call slop now was more generally accepted before AI. Slop is not only the content. It is the volume, and it is the lack of intent or value from the producer. No amount of data is going to fix that.

IAmNeo2 mesi fa

So does the training data only tell the AI what to do and not what 'not' to do? I'm really not versed in the details of training one.