15
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 08 Aug 2023
15 points (100.0% liked)
TechTakes
1437 readers
135 users here now
Big brain tech dude got yet another clueless take over at HackerNews etc? Here's the place to vent. Orange site, VC foolishness, all welcome.
This is not debate club. Unless it’s amusing debate.
For actually-good tech, you want our NotAwfulTech community
founded 1 year ago
MODERATORS
@self @bitofhope I just want to start feeding all the LLMs the Time Cube site and the Dr. Bronner's labels. Except I think the Time Cube guy gets really racist and anti-semitic in there eventually.
so do the LLMs, if you manage to find a part of the corpus RLHF and basic filtering didn’t touch
just the same as when twitlords found out that The Algoriddem had a lot of Special Treatment, it's going to be fun if/when someone leaks the chatgpt prompt (and prompt response filtering/selection) sourcecode
in the meanwhile, I am going to continue being deeply angry every time I run into someone who doesn't understand How Many Design Choices Have Been Made in the deployment and exposure of this heap of turds. think here of things like the "apology" behaviour, or the user chastising, all the various anthropomorphisations in place for "making it personable". some of those conversations have boiled down to "naw bro it's intelligent bro trust me bro you just don't understand" and it absolutely does my head in
There was a burst of submissions about "jailbreaking" ChatGPT, essentially making it output racist stuff. HN was all over that stuff for a while.
there's a fairly active chunk of research in that space. some of the most recent I've seen is llm-attacks.org (which is a riot)
oh chatgpt’s magic is almost entirely just dark patterns. one thing I’d be curious about if source code ever leaked is if the model’s failure cases are being massaged — a bunch of people have started to notice that when GPT enters a failure state, it tends to pull from the parts of its corpus involving religious or sci-fi imagery, which strikes me as yet another manipulative technique among the many that ChatGPT implements to imply there’s something complex happening when there isn’t
I'll have to pay attention to that. usually I just avoid the content because almost all conversations around it make my blood boil
similarly: the accuracy scoring (both per-prompt and general session shit) almost certainly has someone pulling that into revision/avoidance management. which will eventually end up shaping it into something even more hilariously milquetoast
@self Yeah I mean by contemporary standards the Time Cube might be almost benign. I mean I'm not going to reproduce any of the sketchier bits here, but I've seen worse in screenshots from Microsoft Tay.