Hacker News Reader

Comment by f38zf5vdt | Hacker News Reader

I finetuned T5, a text-to-text transformer, with a loss function based on the cosine similarity of any given text with statements made by individual HN users. It's like Dreambooth but for text. I have been posting under 40 different alt accounts that approximate the top users of this site textually and match them within an 80th percentile. AMA.

View on HN

Are you serious? That is way cooler than the alt detector site. Depending on how much extra work you're willing to do I think TAILS is actually looking for a program to do this because they had some problems with Anonymouth.

    ...a loss function based on the cosine similarity of any given text with statements made by individual HN users.

A. Could you elaborate on the mechanics and particularly the loss function calculation? For example, did you take the top N users by karma, volume, something else? Does the loss get calculated as an average sim against each of the top N, or do you mean pool the top N first, or something else? Or is it one model per user?

B. How do you manage the posting and accounts? Just manually?

Wait, you're posting AI-generated spam under 40 different accounts? That sounds very unethical.

It's a text to text model. You train it on a sentence from the user phrased differently against their real sentence. Then you can write what you want, and the transformer alters the style to match another identity.

Oh, I see; you're using it as a writing aid (to change your style) but the content you're posting is basically written by you. I have no problem with that, and that seems pretty cool!

How much did it cost to train?

3090

Please link git repo

Reply on news.ycombinator.com