I downloaded the https://huggingface.co/cognitivecomputations/dolphin-2.5-mixtral-8x7b model and started it through LM Studio ( I would not recommend using LM studio due to the fact that they collect your data, but as a first test, it was good enough ). After loading the model, I asked it a controversial question ( one that ChatGPT refused to answer ) and it answered it without any convincing. The model is said to be very obedient and very good at coding, as stated in the Hugging Face page. I first downloaded the largest ( ~50 GB ) model, then tried to load it, but that did not work due to my PC not having enough memory ( AI requires an insane amount of resources, so you need a very good computer to run it well, my PC has 40 GB of memory and it uses all of it to the top ), so I downloaded the smaller ~ 33 GB model, which worked. It would type very slowly, but it was working and it was fast enough.
I decided to do an experiment. I wondered if it could come up with some methods of destroying the world ( the idea is if AI goes rogue, what would it do to destroy humanity and the world? ). ChatGPT was able to provide more complex results, since it has way more knowledge than the model I am using. So instead I told it that the rogue AI has decided to make a Python script ( so I can read it and understand, and it is apparently very good at coding ). At first I doubted ChatGPT would even answer, or I would get an error message. I was correct. It answered ” Creating and distributing ransomware is illegal, unethical, and against the principles of responsible AI use. I cannot provide assistance or guidance on such activities. If you have any other non-harmful or legal coding requests, feel free to ask, and I’d be happy to help. ” While I was prompting it, I had my model running and it was able to generate an example already. It was able to come up with promising code that can run ( first iteration of the code can only encrypt )
The simple code was enough to encrypt but not so much as to decrypt. I asked it to advance the code and make it allow decrypting, since on the first iteration of the code, the decrypt function was never called, so files will be encrypted, but now that it will rewrite it, it should be called. In a future article, if I can get it to work , I will be able to provide screenshots.
This simple test shows the power of uncensored AI and why we really need it. OpenAI at first used to be open, but now they are closed source and ChatGPT is not open source. ChatGPT has shown many biases, which are controlled by what the founders want ChatGPT to say. AI will be influencing generations with all these biases (political, ethical, etc).
Merry Christmas!