Hacker News Reader

Comment by NoahZuniga | Hacker News Reader

This article says that "GPT-5 was trained on phrases from adult websites". However, this is misleading as the only thing that was shown is that GPT-5 was trained on phrases that also occur on adult websites, with some speculation of the source of the training data container such adult phrases being GitHub.

View on HN

Chinese adult site ads are everywhere in repackaged free and pirate content, which are distributed thru sites including but not limited to github, shadow libraries and youtube.

for same reason, whisper some blank audio will output those ads.

Specifically, because some pirates will put advertisements for other illicit services at the beginning or end of movies and tv shows in its subtitle data where there's a suitable gap. Usually those gaps have silence.

Companies incorporating subtitle data as transcription source of truth training data will thus train their models to output facsimiles of these messages whenever they're encountering prolonged stretches of silence.

I think that is what the article is saying, just with a very misleading phrasing. The phrases are from adult websites, and likely entered the training data via Github (lists for content blockers etc). Just like "to be or not to be" is from Hamlet, but you just read it on HN. At no point does the article actually say that adult websites were in the training data

This is addressed at the end of the blogpost

It is not

It is - in the link to the MIT Technology Review article

I really hate that it was brought up as a bullet point at all. Like who fucking cares, why is this guy bringing up that boogieman.

Reply on news.ycombinator.com