3 ms·

> Releasing completely unhinged LLM models is very problematic Why? > It seems clear that you don't want your end users to be driven into suicide or killing t

> Releasing completely unhinged LLM models is very problematic Why? > It seems clear that you don't want your end users to be driven into suicide or killing themselves accidentally Who are end-users here exactly? This is not a consumer product, the "don't microwave the cat" argument isn't really applicable here.

jonathanstrange 3y ago
> Why? I explained why in the post. > Who are end-users here exactly? Whoever uses it in the end.
- foverzar 3y ago
  > I explained why in the post. No, my friend, you really did not. "end users to be driven into suicide or killing themselves accidentally" is insanely vague and unrealistic. > Whoever uses it in the end. So, people who have enough mental capacity to run them? Is your point actually that we shouldn't release objects or tools because of a vague unrealistic potential threat with an unrealistically small vector? I really can't grasp that logic. How in the world people are thinking that "killing yourself with an LLM" is more likely then an adult dying of putting nails into powersockets or hurting themselves with non-rounded scissors? Killing yourself with LLM is a definite Darwin Award winner. Realistically, something like this can only happen with a mentally handicapped person, up to the point where they need actual supervision, childproofed powersockets and rounded scissors. This is really no more "problematic" than any minor everyday hazard, if not less. Powersockets are everywhere. Launching an LLM actually requires some skills and hardware.
  - jonathanstrange 3y ago
    You've just made up your own arbitrary definitions of "vague unrealistic potential threats" in a self-serving argument - I disagree but also don't think there is much to discuss here. > non-rounded scissors Rounded scissors are standard safety features in first aid kits and many related areas. Couldn't you at least have tried to find a better example for your point? Anyway, this area will be regulated by law very soon whether you like it or not, and removing safety features will not help anyone in future lawsuits. > Powersockets are everywhere They are highly regulated, too. The example I gave of making electric wood burnings by tinkering with home appliances has actually killed a sizeable number of people. In their case, they followed TikTok videos, but they could just as well followed unsafe advice given by an unhinged AI. It's going to happen, I can assure you that. It makes perfect sense to include safety features that stop AIs from giving such advice, and makes perfect sense to stop AI from offending and harassing end users. These issues are relevant in court and it is completely unrealistic to assume the law would follow anyone's "whatever, anything goes" attitude.
    - foverzar 3y ago
      > Rounded scissors are standard safety features in first aid kits and many related areas. Couldn't you at least have tried to find a better example for your point? Isn't "rounded scissors" a slang for safe plastic scissors? I am not a native english speaker, sorry. > Anyway, this area will be regulated by law very soon whether you like it or not and removing safety features will not help anyone in future lawsuits Lawsuits from whom? People who open a box of needles and start plucking their own eyes with them? > They are highly regulated, too. Not in that sense. I can use any kind of regulation or customise it however I like at my own responsibility. I can build by own insane fancy powersocket. And I don't really assume that if I personally chose to use an uninsulated wire for something and it ends up badly, that this s some one else's fault. > In their case, they followed TikTok videos, but they could just as well followed unsafe advice given by an unhinged AI. It's going to happen, I can assure you that. > It makes perfect sense to include safety features that stop AIs from giving such advice, and makes perfect sense to stop AI from offending and harassing end users. I think I see where the confusion is coming from. This is not an AI. You are assuming that this is like a full-blown product, with a specific use case like an advice-giving AI. But it isn't. It's just an LLM. Like, a part, not a mechanism. These restrictions only useful purpose is demonstration of how one could add restrictions in their specific case. They are very generic, don't add any value and have noticeable drawbacks in unforeseen cases. Obviously, when one builds their own AI product with end-user reach and accessibility, they need to build safety features adequate to their usecase, but they might be completely different from what we are imagining.

jonathanstrange 3y ago

> Why? I explained why in the post. > Who are end-users here exactly? Whoever uses it in the end.

foverzar 3y ago
> I explained why in the post. No, my friend, you really did not. "end users to be driven into suicide or killing themselves accidentally" is insanely vague and unrealistic. > Whoever uses it in the end. So, people who have enough mental capacity to run them? Is your point actually that we shouldn't release objects or tools because of a vague unrealistic potential threat with an unrealistically small vector? I really can't grasp that logic. How in the world people are thinking that "killing yourself with an LLM" is more likely then an adult dying of putting nails into powersockets or hurting themselves with non-rounded scissors? Killing yourself with LLM is a definite Darwin Award winner. Realistically, something like this can only happen with a mentally handicapped person, up to the point where they need actual supervision, childproofed powersockets and rounded scissors. This is really no more "problematic" than any minor everyday hazard, if not less. Powersockets are everywhere. Launching an LLM actually requires some skills and hardware.
- jonathanstrange 3y ago
  You've just made up your own arbitrary definitions of "vague unrealistic potential threats" in a self-serving argument - I disagree but also don't think there is much to discuss here. > non-rounded scissors Rounded scissors are standard safety features in first aid kits and many related areas. Couldn't you at least have tried to find a better example for your point? Anyway, this area will be regulated by law very soon whether you like it or not, and removing safety features will not help anyone in future lawsuits. > Powersockets are everywhere They are highly regulated, too. The example I gave of making electric wood burnings by tinkering with home appliances has actually killed a sizeable number of people. In their case, they followed TikTok videos, but they could just as well followed unsafe advice given by an unhinged AI. It's going to happen, I can assure you that. It makes perfect sense to include safety features that stop AIs from giving such advice, and makes perfect sense to stop AI from offending and harassing end users. These issues are relevant in court and it is completely unrealistic to assume the law would follow anyone's "whatever, anything goes" attitude.
  - foverzar 3y ago
    > Rounded scissors are standard safety features in first aid kits and many related areas. Couldn't you at least have tried to find a better example for your point? Isn't "rounded scissors" a slang for safe plastic scissors? I am not a native english speaker, sorry. > Anyway, this area will be regulated by law very soon whether you like it or not and removing safety features will not help anyone in future lawsuits Lawsuits from whom? People who open a box of needles and start plucking their own eyes with them? > They are highly regulated, too. Not in that sense. I can use any kind of regulation or customise it however I like at my own responsibility. I can build by own insane fancy powersocket. And I don't really assume that if I personally chose to use an uninsulated wire for something and it ends up badly, that this s some one else's fault. > In their case, they followed TikTok videos, but they could just as well followed unsafe advice given by an unhinged AI. It's going to happen, I can assure you that. > It makes perfect sense to include safety features that stop AIs from giving such advice, and makes perfect sense to stop AI from offending and harassing end users. I think I see where the confusion is coming from. This is not an AI. You are assuming that this is like a full-blown product, with a specific use case like an advice-giving AI. But it isn't. It's just an LLM. Like, a part, not a mechanism. These restrictions only useful purpose is demonstration of how one could add restrictions in their specific case. They are very generic, don't add any value and have noticeable drawbacks in unforeseen cases. Obviously, when one builds their own AI product with end-user reach and accessibility, they need to build safety features adequate to their usecase, but they might be completely different from what we are imagining.