I read about this earlier on Ars Technica. I was expecting a paywalled link. Was not expecting to find a mention of “No Longer Human.” Ars didn’t mention that. Or the chat logs. It was a long article but didn’t go into the same depth.
So, I’ve read “No Longer Human.” A more recent translation is called “A Shameful Life” and that’s a bit more apt, I think, but doesn’t have the same ring. It’s about a guy who feels less and less like a person, like what he does and feels doesn’t matter. It’s a wild book, about a double suicide, and the author later killed himself much the same way. There have been several adaptations — none of them very good. None of them quite captured the book. I wonder if it’s just unfilmable. Anyway, it’s a shame that it’s being referenced here, because it’s good literature worth considering, and I hate to see it maligned in much the same way as the Doom game was following the Columbine massacre. Relevant or not (guns in that case, suicide in this case), it’s a shame art gets associated with tragedy simply by association.
Perhaps the same could be said of AI technology, and it has been. But certainly AI needs better safeguards. According to Ars, when the guy started asking about suicide, ChatGPT said it could not help him — unless he specified he was talking about fictional characters. So he did that (Ars constantly refers to it as a “jailbreak”) for a while, and then I guess (and they guess as well) that ChatGPT just assumed that context and stopped requiring him to specify that.
Even blenders have safeguards though, if the pitcher isn’t installed most won’t work. I don’t think it’s insane to require some sort of safety with LLMs.
I think the metaphor is finetuning a LLM for ‘safety’ is like trying to engineer the blades to be “finger safe”, when the better approach would be to guard against fingers getting inside an active blender.
Finetuning LLMs to be safe is just not going to work, but building stricter usage structures around them will. Like tools.
This kinda goes against Altman’s assertion that they’re magic crystal balls (in progress), which would pop his bubble he’s holding up. But in the weeds of LLM land, you see a lot more people calling for less censoring, and more sensible and narrow usage.
A blender with glass blades that frequently shatter when at full speed or when cutting hard food. It also has no lid and is marketed as a bath toy.
It’s up to the user to think for themselves and not use it as a bath toy. And to use safety goggles. And check the food for glass shards before each bite.
I read about this earlier on Ars Technica. I was expecting a paywalled link. Was not expecting to find a mention of “No Longer Human.” Ars didn’t mention that. Or the chat logs. It was a long article but didn’t go into the same depth.
So, I’ve read “No Longer Human.” A more recent translation is called “A Shameful Life” and that’s a bit more apt, I think, but doesn’t have the same ring. It’s about a guy who feels less and less like a person, like what he does and feels doesn’t matter. It’s a wild book, about a double suicide, and the author later killed himself much the same way. There have been several adaptations — none of them very good. None of them quite captured the book. I wonder if it’s just unfilmable. Anyway, it’s a shame that it’s being referenced here, because it’s good literature worth considering, and I hate to see it maligned in much the same way as the Doom game was following the Columbine massacre. Relevant or not (guns in that case, suicide in this case), it’s a shame art gets associated with tragedy simply by association.
Perhaps the same could be said of AI technology, and it has been. But certainly AI needs better safeguards. According to Ars, when the guy started asking about suicide, ChatGPT said it could not help him — unless he specified he was talking about fictional characters. So he did that (Ars constantly refers to it as a “jailbreak”) for a while, and then I guess (and they guess as well) that ChatGPT just assumed that context and stopped requiring him to specify that.
No. Much like a kitchen blender, it’s a tool. It’s on the user to not stick their hand into it and turn it on
Even blenders have safeguards though, if the pitcher isn’t installed most won’t work. I don’t think it’s insane to require some sort of safety with LLMs.
I think the metaphor is finetuning a LLM for ‘safety’ is like trying to engineer the blades to be “finger safe”, when the better approach would be to guard against fingers getting inside an active blender.
Finetuning LLMs to be safe is just not going to work, but building stricter usage structures around them will. Like tools.
This kinda goes against Altman’s assertion that they’re magic crystal balls (in progress), which would pop his bubble he’s holding up. But in the weeds of LLM land, you see a lot more people calling for less censoring, and more sensible and narrow usage.
The pitcher doesn’t stop you from sticking your fingers into it if you try, it just makes accidents less likely. Same thing here.
A blender with glass blades that frequently shatter when at full speed or when cutting hard food. It also has no lid and is marketed as a bath toy.
It’s up to the user to think for themselves and not use it as a bath toy. And to use safety goggles. And check the food for glass shards before each bite.