4 ms·

I have been doing this with local LLMs that are "unsafe" using some basic python. I write a diary entry and ask it "Tell me some uncomfortable truths about this

I have been doing this with local LLMs that are "unsafe" using some basic python. I write a diary entry and ask it "Tell me some uncomfortable truths about this person." as the prompt and it roasts me. I then ask it, "Tell me why this person is not on the path to becoming a billionaire" and it says I spend way too much time on my personal life and relationships when I should instead be focused on starting businesses in high-growth industries and networking with high net worth individuals. Some of the more standard LLMs try very very hard not to insult me and tell me to talk to a licensed professional for everything. That's not what I want. I wanted to get the most impolite LLMs I could, so I looked up some list about which ones were the best for topics like erotic roleplay and those worked way better for even non-sex topics because they're not RLHF'd to be unconditionally nice or excessively cautious.

dartos 3y ago
Having a robot roast you daily because of troubles you have doesn’t sound like therapy. It sounds like a bad relationship and you should break up with them.
- narrator 3y ago
  This is a good point. However, how do I get better feedback than "everything you do and feel is ok as long as it's not criminal," and "improve your diet, sleep, exercise and consult with a professional" for everything else? I want creative insights into my life that I don't already know. Putting in "Why am I not a billionaire yet?" is a lofty enough goal that I am not beating myself up too badly for not meeting that goal. "Why am I not a world class athlete for my age group?" could be another one.
  - dartos 3y ago
    > However, how do I get better feedback than "everything you do and feel is ok as long as it's not criminal," and "improve your diet, sleep, exercise and consult with a professional" for everything else? Knowing and trusting people who care about you. Building a support system. Asking them. Going to a support group (online or not) Not by detaching yourself from reality and asking self help questions to a statistical number generator probably made by some of the largest tech companies around. (With a consistent track record of causing and profiting off mental health problems and addictive dark UX patterns)
laborcontract 3y ago
For code, I've tried to use Claude to give me condescending, highly opinionated answers and give me brutal feedback in a way that a toxic coworker/senior would do. I didn't think it would do it, but it did, and it actually ended up making me feel really crummy about myself. Had to turn it off after 5 iterations.

dartos 3y ago

Having a robot roast you daily because of troubles you have doesn’t sound like therapy. It sounds like a bad relationship and you should break up with them.

narrator 3y ago
This is a good point. However, how do I get better feedback than "everything you do and feel is ok as long as it's not criminal," and "improve your diet, sleep, exercise and consult with a professional" for everything else? I want creative insights into my life that I don't already know. Putting in "Why am I not a billionaire yet?" is a lofty enough goal that I am not beating myself up too badly for not meeting that goal. "Why am I not a world class athlete for my age group?" could be another one.
- dartos 3y ago
  > However, how do I get better feedback than "everything you do and feel is ok as long as it's not criminal," and "improve your diet, sleep, exercise and consult with a professional" for everything else? Knowing and trusting people who care about you. Building a support system. Asking them. Going to a support group (online or not) Not by detaching yourself from reality and asking self help questions to a statistical number generator probably made by some of the largest tech companies around. (With a consistent track record of causing and profiting off mental health problems and addictive dark UX patterns)

laborcontract 3y ago

For code, I've tried to use Claude to give me condescending, highly opinionated answers and give me brutal feedback in a way that a toxic coworker/senior would do. I didn't think it would do it, but it did, and it actually ended up making me feel really crummy about myself. Had to turn it off after 5 iterations.