@mttaggart Oh God. Tell me it's bots and trolling.

It's like this: between all examples of, say, a contract in the training corpora, what would you imagine the variance is between them? Consider variance as diction, syntax, and even conceptual structure. Now consider the variance between all examples of, like, a login process. It's so, so much harder to distill not-code text to produce useful output in similar shapes.But of course, the insecure nature of the example code also guarantees a tendency to insecurity in model output as well.

@tomcatHot take: if digital privacy is as unattainable as the people telling us to give up on it was, none of this would be necessary.@mttaggart

@mttaggart as much as i thoroughly detest every link in the chain that made this Ars article possible--from the asshat who made the agent which submitted the PR to the asshats at Ars--question 3 in particular still made me gasp involuntarily. It's such a great line of questioning. Reframing AI hallucinations like this sometimes feels like the only way to make fence-sitters come down. "If you had a coworker who was completely wrong 10% of the time when you asked them a question... would you ask them questions anymore?" has been how I've gotten through to a few people in my life.

@mttaggart I feel like "the author in question won’t work with ars anymore" would have been a better answer, tbh. Yes this might happen, but really…

@mahryekuh @mttaggart Just look up your ID with @haveibeenpwned to see if you exist or not

@mttaggart @scottwilson thank you for such a good post summarizing the Anthropic study - I was trying to share the paper with some folks earlier, but they aren’t used to reading academic docs, so your post has been a much easier distillation to send them

@mttaggart “aye… I could do that…”

@da_667 @mttaggart *unless* they find a new sucker.... I mean investor. Or they bribe the orange... I mean get public bailout.