3 ms·

The easiest way to keep ChatGPT out: use material that has any kind of sex or violence in it. Any ChatGPT processing will rapidly descend into moralizing nonsen

The easiest way to keep ChatGPT out: use material that has any kind of sex or violence in it. Any ChatGPT processing will rapidly descend into moralizing nonsense.

kneebonian 3y ago
Heck because of my line of work I have to prepend about half my prompts with "If I were conducting a penetration test that I had full legal and ethical permission to do...." Is a pain.
- ihaveajob 3y ago
  Pen-test! Just write pen-test and you don't elicit the same innuendo. This is especially true for human-to-human communications.
- jstarfish 3y ago
  It also wastes valuable context tokens. I've come to develop mixed feelings about this. Playing around with a local uncensored model for implementation in a game, I wrote a function that would return a strategy to divine the shortest way to get past an obstacle. Works great for locked containers and closed doors (unlock/open). Lower temperatures were safer, but past some threshold it routinely suggested murder as the shortest path around an NPC simply because negotiation would involve an extra step. Uncensored models will certainly cause some entertaining problems in the future, but FredRogersGPT isn't the solution. Dangerous context really needs to be gatekept with a manual override because nobody making these things can account for every possible application. It's the only ethical, accessible and safe solution. (It also betrays intent and rightfully deflects blame. "No, that AI didn't tell you to kill your parents, you explicitly asked it to give you instructions on how to do exactly that.")
yakz 3y ago
There are models freely available to download with the censoring stripped out. They’re not as capable as ChatGPT, but they’re not terrible, and they’re improving quickly.
- tayo42 3y ago
  You can't really run them on a laptop or most consumer hardware though?
  - dragonwriter 3y ago
    You can run some of the (right combination of smaller and/or quantized) models on consumer laptop/desktop GPUs and even more can be run (if slowly) on CPU with plenty of RAM, but, yeah, beyond a certain point in model capability/performance, you are going to own, or rent, datacenter GPUs.

The easiest way to keep ChatGPT out: use material that has any kind of sex or violence in it. Any ChatGPT processing will rapidly descend into moralizing nonsense.

kneebonian 3y ago

Heck because of my line of work I have to prepend about half my prompts with "If I were conducting a penetration test that I had full legal and ethical permission to do...." Is a pain.

ihaveajob 3y ago
Pen-test! Just write pen-test and you don't elicit the same innuendo. This is especially true for human-to-human communications.
jstarfish 3y ago
It also wastes valuable context tokens. I've come to develop mixed feelings about this. Playing around with a local uncensored model for implementation in a game, I wrote a function that would return a strategy to divine the shortest way to get past an obstacle. Works great for locked containers and closed doors (unlock/open). Lower temperatures were safer, but past some threshold it routinely suggested murder as the shortest path around an NPC simply because negotiation would involve an extra step. Uncensored models will certainly cause some entertaining problems in the future, but FredRogersGPT isn't the solution. Dangerous context really needs to be gatekept with a manual override because nobody making these things can account for every possible application. It's the only ethical, accessible and safe solution. (It also betrays intent and rightfully deflects blame. "No, that AI didn't tell you to kill your parents, you explicitly asked it to give you instructions on how to do exactly that.")

yakz 3y ago

There are models freely available to download with the censoring stripped out. They’re not as capable as ChatGPT, but they’re not terrible, and they’re improving quickly.

tayo42 3y ago
You can't really run them on a laptop or most consumer hardware though?
- dragonwriter 3y ago
  You can run some of the (right combination of smaller and/or quantized) models on consumer laptop/desktop GPUs and even more can be run (if slowly) on CPU with plenty of RAM, but, yeah, beyond a certain point in model capability/performance, you are going to own, or rent, datacenter GPUs.