Hacker News Reader

Comment by astrange | Hacker News Reader

Anthropic says

> To date we have not used any customer or user-submitted data to train our generative models.

https://www.anthropic.com/news/claude-3-5-sonnet

There's an obvious problem with the concept of training on user prompts; how would training on a bunch of questions cause it to know the answers?

View on HN

"There's an obvious problem with the concept of training on user prompts; how would training on a bunch of questions cause it to know the answers?"

I imagine by analysing the chat? If the user says thanks in the end, or gives a thumps up, it likely was a useful and correct answer, that could be included in further training. Or at least considered for future training and I cannot imagine them not considering and experimenting with it.

User queries were at least historically useful to train smaller models from larger models. You need to know the kind of questions real people ask to train a model that’s good at answering those questions

Back when I started using LLMs for writing code I would type out long, gently phrased explanations about why it was wrong, as if I was teaching a pupil, hoping it would help. I'm sure a lot of us did. If they can parse and mine those prompts, they'll have a nice little metacorpus to build on.

Now I just tell it to stop being stupid over and over until it does a good job. I wonder if it would improve the model to keep all of the beratement in the training data.

Edit: Apparently a 'metacorpus' is a swollen nematode ass. My sincerest apologies, bros.

Reply on news.ycombinator.com