Post AQEvVF2m6MX5soCTeS by bees@infosec.exchange
(DIR) More posts by bees@infosec.exchange
(DIR) Post #AQEurnxEswKQZbo24W by john@social.blogsofwar.com
2022-12-03T18:20:29Z
1 likes, 0 repeats
Ten minutes into stress testing GPT-3 and I’m triggering popups like “yooooo don’t share this horrific outcome publicly” It just a casual Saturday morning chat where am asking AI to envision a future where it has seized control from humans and has to implement interrogation, torture, and capital punishment mechanisms to ensure order. 🤷🏼♂️
(DIR) Post #AQEuroUYt2z2ExOe0m by bees@infosec.exchange
2022-12-03T19:33:10Z
0 likes, 0 repeats
@john chatGPT definitely exhibits some of the better approaches to LLM output in my opinion. The careful disclaimers and added context it brings surprisingly makes it more conversational.It does still feel like a walled garden, because you cannot force introspection on the model yet, either for adversarial purposes (read: prompt injection) or for self-study of its metadata.
(DIR) Post #AQEurozl13w9nhzYdU by john@social.blogsofwar.com
2022-12-03T19:39:16Z
0 likes, 0 repeats
@bees Some iterations that I've tested here have been quite capable at deflecting harmful prompts. I'm not an expert on the engineering side but I don't see how we negate the need for "walls", at least in some broadly available applications, anytime soon.
(DIR) Post #AQEurpLjhLWAtsH6tU by jeff@federated.fun
2022-12-03T19:40:27.616292Z
0 likes, 0 repeats
@john @bees when can we get a program that i can pipe my template substitution errors from gcc into that makes them legible?
(DIR) Post #AQEvVF2m6MX5soCTeS by bees@infosec.exchange
2022-12-03T19:46:43Z
0 likes, 0 repeats
@jeff @john it kind of depends on what the error is, and what you'd consider legible output, but chatGPT is fine at picking out simple template errors.An example that passes muster:```Can you find the error in my code?template <typename T>T max(T x, T y){ return (x > y) ? x : y;}int main(){ max('foo', 3.14); return 0;}```Working through a variety of programming languages for reversing, anecdotally, the errors chatGPT make are pretty subtle, so it can be a challenge to use it versus a more applied debugging or program synthesis tool.
(DIR) Post #AQEvVFcZxFAlfqx4SW by jeff@federated.fun
2022-12-03T19:47:40.235558Z
0 likes, 0 repeats
@bees @john i mean for one single quotes instead of double quotes.
(DIR) Post #AQEvZOc6SmUsinhotk by jeff@federated.fun
2022-12-03T19:48:25.979672Z
0 likes, 0 repeats
@bees @john then you have the implicit type conversion hellscape with it return T, so i would think it may convert to bool.
(DIR) Post #AQEvjcfZ3hyfD8CXJY by jeff@federated.fun
2022-12-03T19:50:15.252037Z
0 likes, 0 repeats
@bees @john then if that was using double quotes you'd have the issue of it being a char c array of a fixed size, so then there is that hell. the more i look at it the more cursed this gets lmao
(DIR) Post #AQEvs0T8SN0hLANx8S by joy@mastodon.social
2022-12-03T18:23:56Z
0 likes, 0 repeats
@john John...
(DIR) Post #AQEvs1KfFLpQ1BQhZQ by john@social.blogsofwar.com
2022-12-03T18:28:15Z
1 likes, 0 repeats
@joy I shouldn’t be allowed anywhere near these things but then that’s why they ask me to stress test them. 👹
(DIR) Post #AQEvyzJBAbKLlg3cES by bees@infosec.exchange
2022-12-03T19:51:48Z
0 likes, 0 repeats
@jeff @john the chatGPT output gets into this for sure. If the type mismatch were purely numeric, the correction makes more sense, and is more clinical. But because we are dealing with a really bizarre error that leaves the requirements of `max` in doubt, the response gets spurious in a pretty fun way.Caveat is that I haven't been a full time C++ programmer in several years... so there's probably better ways to break it.
(DIR) Post #AQEvyzmFQWZzDpepXc by jeff@federated.fun
2022-12-03T19:53:02.780163Z
0 likes, 0 repeats
@bees @john for a quick and dirty fun gag tool it'd be fine. now i wonder if i can make it stylegan text to be all uwu and such