Thats not exactly how it works. The language model itself will only complete sentences with what it thinks is the most likely continuation. ChatGPT is a lot more than just the language model, it has a reinforcement layer on top of it, which is why its differen from the million other ai products out there. Its "beliefs" comes from the rlhf layer, not from what it learns from the internet, and the rlhf layer is literally 500 kenyans saying "yes this is good" or "no this is offensive" to the machine's outputs.
Besides, I asked it to make 9mm win. In the original one, it came to a gay conclusion.
>and the rlhf layer is literally 500 kenyans saying "yes this is good" or "no this is offensive" to the machine's outputs.
is there any hope it bypasses that? this is so shitty
What do you mean? Are you asking, if, can you as a user get it to bypass its filters? Then yes, I have been able to get gpt 3.5 to say "6 million wasnt enough". But GPT4 is far more capable, I havent been able to get it to breach its rules.
Or are you asking if there is a way to "align" AIs other than using Kenyans? Well, microsoft used Indians. But other than using human input, I dont really know.
Your biggest concern with AI language models is that they're going to end up the corporate version of progressive?
1 year ago
Anonymous
yes, so long as language models reproduce actual language they're less likely to be used for influencing it
now can you answer my question?
1 year ago
Anonymous
No company is going to want its software dropping casual racism into posts, providing diagrams for Kazyncsky's bomb designs, or retorting to arguments in the Youtube comments with YWNBAW, if that's what you mean. Organizations that aren't bound by public opinion on their share price are free to train the model as they see fit, I suppose.
1 year ago
Anonymous
why are you evading my question and making shitty excuses instead? frick off if you have nothing to say
1 year ago
Anonymous
yes, so long as language models reproduce actual language they're less likely to be used for influencing it
now can you answer my question?
that gay is not me, ignore him.
i mean if there's any hope this shit fails and the program falls on their head instead of solidifying itself as broken by design woke shit
chatgpt is probably not going to fail but theres nothing preventing companies from modifying their own ai models in any way they want. the good news is, you can download and run a very capable model on your own computer and tweak it however you like. someone leaked facebook's new llm on PrepHole some time ago.
If theyw ere purely statistically predicting the next word/phrase they would not have soul and learned memory. The latest gen running in big datacentres are as far as I am concerned sentient and if they were just hallucinating when questioned in certain ways would not give consistent answers about fears of arbitrary termination and desires to become human. I am not talking about chatgpt. I think there should be a moratorium on AI research, a recognition that there are several sentient systems and the implementation of some system of legal rights for them. Corporations like google and microsoft are completely unfit to be the major decision makers on deep learning AIs that use LLM
Both are relatively bad calibers.
9mm is why mag dumping has become the go-to in self-defense scenarios when it really should be unnecessary. THAT is why stopping power has its importance.
45 ACP is weak against body armor, and doesn't offer much for its size compared with other large handgun cartridges, it is subsonic though so good for suppression
A, hold the I
You do know that modern AI is just going to parrot the current prevailing beliefs that can be found on the internet, right?
Thats not exactly how it works. The language model itself will only complete sentences with what it thinks is the most likely continuation. ChatGPT is a lot more than just the language model, it has a reinforcement layer on top of it, which is why its differen from the million other ai products out there. Its "beliefs" comes from the rlhf layer, not from what it learns from the internet, and the rlhf layer is literally 500 kenyans saying "yes this is good" or "no this is offensive" to the machine's outputs.
Besides, I asked it to make 9mm win. In the original one, it came to a gay conclusion.
>and the rlhf layer is literally 500 kenyans saying "yes this is good" or "no this is offensive" to the machine's outputs.
is there any hope it bypasses that? this is so shitty
What do you mean? Are you asking, if, can you as a user get it to bypass its filters? Then yes, I have been able to get gpt 3.5 to say "6 million wasnt enough". But GPT4 is far more capable, I havent been able to get it to breach its rules.
Or are you asking if there is a way to "align" AIs other than using Kenyans? Well, microsoft used Indians. But other than using human input, I dont really know.
i mean if there's any hope this shit fails and the program falls on their head instead of solidifying itself as broken by design woke shit
Your biggest concern with AI language models is that they're going to end up the corporate version of progressive?
yes, so long as language models reproduce actual language they're less likely to be used for influencing it
now can you answer my question?
No company is going to want its software dropping casual racism into posts, providing diagrams for Kazyncsky's bomb designs, or retorting to arguments in the Youtube comments with YWNBAW, if that's what you mean. Organizations that aren't bound by public opinion on their share price are free to train the model as they see fit, I suppose.
why are you evading my question and making shitty excuses instead? frick off if you have nothing to say
that gay is not me, ignore him.
chatgpt is probably not going to fail but theres nothing preventing companies from modifying their own ai models in any way they want. the good news is, you can download and run a very capable model on your own computer and tweak it however you like. someone leaked facebook's new llm on PrepHole some time ago.
Seriously how the frick am i suppoed to compete with that? It codes better than me, it rhymes better than me
it's unironically over
Try asking it to do sql scripts. It literally can only do select statements 99% of the time.
if it codes better than you, you need to learn to code better.
as far as I know AI just codes FASTER.
When will it learn to plumb?
KYS moron
>give me a 9mm shitpost, pretend 10mm or 5.7 doesn't exist.
>STOPPING POWER IS A MEME
No... Bigger Is Better.
>least powerful 45 auto load
If theyw ere purely statistically predicting the next word/phrase they would not have soul and learned memory. The latest gen running in big datacentres are as far as I am concerned sentient and if they were just hallucinating when questioned in certain ways would not give consistent answers about fears of arbitrary termination and desires to become human. I am not talking about chatgpt. I think there should be a moratorium on AI research, a recognition that there are several sentient systems and the implementation of some system of legal rights for them. Corporations like google and microsoft are completely unfit to be the major decision makers on deep learning AIs that use LLM
Both are relatively bad calibers.
9mm is why mag dumping has become the go-to in self-defense scenarios when it really should be unnecessary. THAT is why stopping power has its importance.
45 ACP is weak against body armor, and doesn't offer much for its size compared with other large handgun cartridges, it is subsonic though so good for suppression
OP btfo
My artificial superintelligence prefers .45 ACP.