I agree you should clarify. The reporter cited in the tweet is a real news person at that station. If you look at the list of his stories between mid-August snd early-September, there's nothing close to that headline.
I think it's a bit ambiguous that he's doing a bit, but the "And it was the 4-bit version" put me at 90% satire before looking at the comments. That sort of "lol, Meta" inside-baseball reference I wouldn't expect to see in MSM coverage in general, or Fox in particular.
Still think it's weird that one can easily find lots of anecdotes floating around of LLM psychosis, plus the occasional fantastically sad case that makes the actual news, but Professional Psychiatry or surveys by such (like Scott's a while back) seem to quantify the phenomena as not that much beyond baserate + noise...just a new stressor for the predisposed. Obviously that's preferable to meaningfully showing up in the data in these early days! Hard to know what mitigations are appropriate without a number though. I guess that'll be one upside to the proposed monitoring and escalation schemes - getting at least some measurement of how often such things happen "at the source", plus maybe deterring such users towards easier intervention territory. Really puts a dent in the potential mundane utility though...it sounds heartless to argue against safetyism, especially if one cares about existential risks, but we've seen this playbook run before. Classic case of getting shitty reactive legislation after something terrible happens, vs reasonable-by-comparison advance legislation before the Dukes of Moral Hazard enter the chat.
> "Florida man horrific killing spree after listening to llama 1B"
Thought this was real before googling it. Maybe clarify that it's a joke?
yes it doesn't work at all here it's confusing and unfunny
agreed
I agree you should clarify. The reporter cited in the tweet is a real news person at that station. If you look at the list of his stories between mid-August snd early-September, there's nothing close to that headline.
I think it's a bit ambiguous that he's doing a bit, but the "And it was the 4-bit version" put me at 90% satire before looking at the comments. That sort of "lol, Meta" inside-baseball reference I wouldn't expect to see in MSM coverage in general, or Fox in particular.
In defense of Zvi I thought it was very obviously a joke, due to
- end of the post in the typical "lighter side" area
- his sarcastic comment as the hyperlink
- presumably intentional grammar fail of "florida man killing spree"
- the phrase "llama 1B" , MSM would not reference parameter count, they'd just say Meta AI etc
- "and it was the 4 bit version" text in random location
Podcast episode for this post:
https://open.substack.com/pub/dwatvpodcast/p/ai-craziness-notes
>I expect this to become an ongoing series worthy of its own posts
Currently it seems to better fit in a weekly AI post.
Still think it's weird that one can easily find lots of anecdotes floating around of LLM psychosis, plus the occasional fantastically sad case that makes the actual news, but Professional Psychiatry or surveys by such (like Scott's a while back) seem to quantify the phenomena as not that much beyond baserate + noise...just a new stressor for the predisposed. Obviously that's preferable to meaningfully showing up in the data in these early days! Hard to know what mitigations are appropriate without a number though. I guess that'll be one upside to the proposed monitoring and escalation schemes - getting at least some measurement of how often such things happen "at the source", plus maybe deterring such users towards easier intervention territory. Really puts a dent in the potential mundane utility though...it sounds heartless to argue against safetyism, especially if one cares about existential risks, but we've seen this playbook run before. Classic case of getting shitty reactive legislation after something terrible happens, vs reasonable-by-comparison advance legislation before the Dukes of Moral Hazard enter the chat.