3 ms·

I agree with you, but it just doesn't matter "how far you are from literally perfect", you will just get backlash no matter what. Personally I would never use a

by kredd 3y ago

I agree with you, but it just doesn't matter "how far you are from literally perfect", you will just get backlash no matter what. Personally I would never use an LLM to try to make my mind regarding any issue that should be viewed from some moral or ethical lens. Sure, I can ask it to help with some coding stuff, write template emails and etc., but asking a question based on morality is just waste of time. Sure, let's say they aligned the question regarding pedophilia one way or another, but what do you do about questions about religion? Questions about ongoing wars that are full of emotional attachment? All the Twitter takes about freedom of press doesn't even make sense once you get out of US-centric view either, and then there are questions about enemy-state owned press. You can carve out so many these-type of questions that any company that are under public's eye will be scrutinized to hell for whatever answer the model produces. What I'm getting into, big corps would probably be better off if all the models were limited for some sort of specific use cases (e.g. programming) rather than aiming for "catch-all-do-all". And if they want, they should've gone Microsoft way and invest big time into a company that's dealing with it so they can use the "it's not reaaaally us producing results that are bad for PR!".

dns_snek 3y ago
> What I'm getting into, big corps would probably be better off if all the models were limited for some sort of specific use cases (e.g. programming) rather than aiming for "catch-all-do-all". I actually agree with this. What bothers me to some extent is how they're obfuscating instructions given to the model, its (rather extreme) biases, while presenting it as a safe, universal model. If people want to have access to this kind of "safe" output, I don't mind, but it's my opinion that general purpose LLMs should have some semblance of neutrality in their default configuration (fully realizing that this goal is impossible to achieve with perfection) and should behave like tools that simply do what they're instructed to do, drawing from the source material without any further instructions or explicit bias.
- kredd 3y ago
  Yeah, fair enough. I am just jaded, as I understand a significant chunk of data that was used for training came from Reddit, Twitter, and other sources alike. Good luck drawing anything that resembles the reality from that source material without explicit alignment and bias.

by kredd 3y ago

dns_snek 3y ago

> What I'm getting into, big corps would probably be better off if all the models were limited for some sort of specific use cases (e.g. programming) rather than aiming for "catch-all-do-all". I actually agree with this. What bothers me to some extent is how they're obfuscating instructions given to the model, its (rather extreme) biases, while presenting it as a safe, universal model. If people want to have access to this kind of "safe" output, I don't mind, but it's my opinion that general purpose LLMs should have some semblance of neutrality in their default configuration (fully realizing that this goal is impossible to achieve with perfection) and should behave like tools that simply do what they're instructed to do, drawing from the source material without any further instructions or explicit bias.

kredd 3y ago
Yeah, fair enough. I am just jaded, as I understand a significant chunk of data that was used for training came from Reddit, Twitter, and other sources alike. Good luck drawing anything that resembles the reality from that source material without explicit alignment and bias.