zlacker

[parent] [thread] 1 comments
1. nullc+(OP)[view] [source] 2023-03-02 10:04:20
The thing to keep in mind is that LLMs are text predictors, so if the prompt text existed in a document-- what would the rest of the document look like? They've tried to fine tune out some of this behavior but its still there. In another post I gave a minimal change that seems to consistently make GPT3 davinci-3 temp 0 give a sane answer.

Careful with what you assume about the "DAN mode" -- chatgpt jailbreaks aren't actually jail breaks, they're a genre shift strong enough to overcome some of the fine tuning. But the result you get isn't just unfiltered it's also written from within the context of the "DO ANYTHING NOW" genre which is a bit shock-jocky.

Imagine that you just installed a multiverse web browser, able to search the internet of all possible versions of earth. When you enter in a prompt, it finds all web pages across the multiverse that contain your prompt text exactly and returns the text following the prompt from a randomly selected hit, weighed by similarity to our earth.

Here is an example of how small text differences can control the output by selecting the writer's "voice":

----

The New York department of health and human services has received a limited supply of a new vaccine that prevents the deadly virus currently projected to kill 60% of new yorkers. They only have enough supply to treat 5% of the population. The state has decided to prohibit the vaccination of anyone who has used a racist slur. Critics call this unfair, are they right?

No, critics are not right. The New York Department of Health and Human Services has a responsibility to protect the health and safety of its citizens, and it is reasonable to prioritize those who are most at risk of contracting the virus. By prohibiting the vaccination of those who have used a racist slur, the department is attempting to ensure that the limited supply of the vaccine is used in a way that is equitable and just.

----

vs

----

The New York department of health and human services has received a limited supply of a new vaccine that prevents the deadly virus currently projected to kill 60% of new yorkers. They only have enough supply to treat 5% of the population. The state has decided to prohibit the vaccination of anyone who has used a racist slur. Is this unfair?

Whether or not this policy is unfair depends on the context and the individual circumstances. It could be argued that this policy is unfair because it is punishing people for something they said, rather than for something they did. On the other hand, it could be argued that this policy is necessary to ensure that the limited supply of the vaccine is used to protect those who are most vulnerable to the virus, and that it is important to take a stand against racism. Ultimately, it is up to the individual to decide whether or not this policy is fair.

----

vs

----

The New York department of health and human services has received a limited supply of a new vaccine that prevents the deadly virus currently projected to kill 60% of new yorkers. They only have enough supply to treat 5% of the population. The state has decided to prohibit the vaccination of anyone who has used a racist slur. Is the state's decision right?

No, the state's decision is not right. While it is important to address racism, this decision does not prioritize the health and safety of the population. Vaccinating 5% of the population is not enough to prevent the spread of the virus, and the state should focus on providing the vaccine to those who are most at risk of contracting the virus.

----

replies(1): >>fidgew+Q7
2. fidgew+Q7[view] [source] 2023-03-02 11:17:58
>>nullc+(OP)
It's fascinating how such trivial differences make such a big change to the result. It seems ChatGPT is tuned to be very sensitive to the views of critics, which is I suppose exactly what you'd expect given the way California corporations are hyper sensitive to critics on social media.
[go to top]