making me feel despair lately: the realization that sharing information with strangers is pretty much always going to lead to that information being included in industrial LLM training data (absent extreme levels of security discipline). even if we ran a usb key sneakernet, all it takes is one person plugging the drive in on a computer with networked backups. even if we distributed everything on paper, all it takes is one mobile phone camera app that stores its data in the cloud
1 Boost
0 Likes
@tiotasram @aparrish poisoning is a valid strategy but I find it too unreliable. LLMs already spout obviously incorrect info all the time and people use it anyway
so many workflows for basic tasks run through "the cloud," and i feel like even technically inclined and otherwise security literate computer users nowadays don't really have a good grasp on which EULA is applying to what service, or what data permissions they've opted into/out of. and also cloud providers are extremely incentivized to train on your data without real consent, and then justify it to you later by saying it was "anonymized" or whatever
@aparrish Yeah my zine dream also involves a fully analog production pipeline, writing on typewriters, film photography, and duplicating on a risograph (technically digital, so I might go even lower tech later). I don't trust any computer connected to the Internet, and I haven't had the time to get into retrocomputing for computers that were designed to operate without the Internet (modern computers feel hamstrung offline).
Of course, I also haven't had the chance to learn to use risographs.
@aparrish I think there is value to delaying the data reaching LLMs, even if that's not as satisfying as denying it to them entirely. If humans are able to understand and act on information/art before the chatbots begin regurgitating it, I'll count that as a win.
I have a vague plan to distribute paper zines that ask the reader to please avoid digitizing them until some time has passed, perhaps a year.
@skyfaller i think you're probably right, and I like this strategy. but i also think the strategy only works if your audience consists entirely of people who understand that "digitization" includes things like... taking a photo of a page and texting it to a friend. which feels like a very small audience
@aparrish valid, but to me there's an opportunity to be had, in a way, since this also means that anyone can poison the models. The poison power of one person is diluted of course, but come up with specific-enough word combinations and there's a decent chance you'll be able to key them up later from a model, I think.
If "everyone develop your own weird works of writing and make up fake facts to put on a fake facts page" became the norm, I think it would be cool (as long as the fake facts are clearly labeled as such where linked to, abusing the fact that scrapers won't carry content across links).
I'm also hoping that the 2027 crash (which I can't manifest hard enough) will dramatically change the equations though.