logoalt Hacker News

chrismorgantoday at 8:05 AM2 repliesview on HN

No, only like any other closed source software or dependency. It’s at the very least harder to hide things like that if all the source is available.

The title here is a misnomer; none of these models are open source, they’re only what is most commonly called open weights (though I’m still not comfortable with the word “open” in there, rather than “free” like in “freeware”).


Replies

thesztoday at 8:30 AM

  > It’s at the very least harder to hide things like that if all the source is available.
In the training of LLMs, the order of batches and their content can introduce adversarial behavior [1].

[1] https://www.pure.ed.ac.uk/ws/portalfiles/portal/256761768/Ma...

You can get all the source to validate training and weights and still end up with adversarial behavior in the model.

dgellowtoday at 9:35 AM

It’s 2 different concepts. Some models are indeed open source, while others are only open weight. But you’re right they are talking about the whole category of open models, which includes both open weight and open source

show 1 reply