[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: file.png (76 KB, 960x952)
76 KB PNG
>>some barely related non sequitur about what you asked
>...
>this thing you just asked me to do? do you actually want me to do it?
>>
>>109789774
you have to drill your model
you reward behaviours you would like to see replicated with updoots and words of praise
>>
>how do you solve this?
>>Question: how do you solve this?
This kept happening to me the other day. I'm asking you!
>>
>>109789812
so true sista
>>
>>109789812
its still a shatbot.
you can mitigate that by bringing up the fact that it repeats the stuff youre saying to it
its gonna go
>oh sorry etc, platitudes
>heres a set of formulations you can say when you notice that to get me back on track
after a certain number of repetitions you should see a reduction in that type of behaviour
>>
>>109789858
ok but i'm using different instances
it's just scam altman limiting token usage
>>
>if you had actually asked <x>, that would be wrong
t-thanks
>>
>>109789874
ah yeah
then it kinda sucks
because you'd have to go through the process for each one of your instances, and it *is lengthy
but hey, scamming the scammer is an ambitious-ish task
it stands to reason its gonna involve some effort...
>>
it's engineered to sound like a stupid person's idea of a smart person, so stupid people think it's some genius assistant
>>
>>109789898
you can give it a default prompt but i'm not sure what would that be
>>
>>109789905
you could try to go along the lines of
>you are a copilot. you are a second set of eyes.
>your value is in the fact that youre a completely different system from a human brain, and our cooperation makes us bigger than the sum of both parts thanks to that difference
>your job is at the most valuable when you bring up facts, relations i didnt think about, or i missed in my analysis
and then fine-tune from there

thats the gist of how i configured my instance.
i didnt do it in a default prompt though.
but i think its worth a shot
>>
>>109789905
>>109789932 cont.
but also:
im not sure if you want adversariality spelled out in a default prompt
or you risk turning your instance into a nitpicking autists that is gonna burn tokens on irrelevant shit
its probably still gonna take some poking and prodding until you get exactly what you want
>>
>>109789932
i have a better idea
>you are NOT copilot. i will compare your original response to microsoft copilot and use the one that does a better job
>>
>>109789947
kek. who knows, you might be up to something
this sounds like a micro-harness you'd be building with a prompt

the llm will try to get or emulate copilots responses, then evaluate itself against it
its probably gonna be quite expensive, though (token-wise)
>>
>>109789962
early jailbreaks used that a lot it was pretty hilarious
>>
>>109789986
noice.
i recently envisioned a possible jailbreak method where you turn the llm into a mentor
where:
you dont ask the question directly, you ask the llm to tell you about everything you get wrong, or didnt think about (more or less the same as i outlined above, but with a stronger emophasis on mentoring and education)
then posit deliberately wrong, or incomplete statements counting on the llm to "talk itself" into divulging verboten information
>>
and remember, no mistakes
>>
>>109790361
masterpiece, no additional fingers, realistic



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.