Almost everyone uses ChatGPT or some form of Large Language Model (LLM) these days. Most people will probably have tried to see what kind of questions LLMs will and will not answer.
Many commercial LLMs have all kinds of safeguards built in that prevent it from answering questions that may be immoral, illegal or otherwise inappropriate.
Recently I heard that you can also run LLMs on your local computer (obviously depending on its performance) and that this opens the use for uncensored models! These are models that do not have those failsafes built in.
Using ollama I managed to install two versions of the llama2 model on my laptop for comparison to how it would respond to certain questions. See some of the results below:
As you can that is quite some different output. Here is another:
It is interesting to see that the censored model always assumes that you are going to do something wrong with the information and not consider curiosity as being the reason behind the question.
Asking a not-so naughty question also gives different results:
I will try and see what more differences I can spot in the meantime!