Do you have a dumb question that you're kind of embarrassed to ask in the main thread? Is there something you're just not sure about?
This is your opportunity to ask questions. No question too simple or too silly.
Culture war topics are accepted, and proposals for a better intro post are appreciated.
Jump in the discussion.
No email address required.
Notes -
(Mildly) Interesting observation:
Recently, people on Twitter have claimed that Claude 3.5 Sonnet is stunningly good at guessing a user's ethnic background from any substantive amount of example text, even examples which have no glaringly obvious tells.
I decided to randomly throw in 2 to 3 comments >500 words each in two tries to see how it did.
In both cases, it correctly guessed I was Indian. The first try had a comment which tangentially mentioned the fact that I was a doctor who had relocated to the UK for training and some tendency to use British spelling, which immediately made it jump to South Asian/Indian. I decided to omit that one and go off more esoteric comments, and once again it got it bang on. I won't share the full chat since it would be around 90% my own comments copied and pasted, but for the last, stunning, example, Claude noticed:
I'm blown away. I had no idea that you could even make that kind of derivation, none of these strike me as Indian™ tropes in the least. All LLMs are excellent world modelers (and by extension at modeling the user), but that's better than I expected and by far.
I'd be curious if anyone else wants to give it a try and is willing to report back. Just copy two or three substantive comments and throw 'em in the pot.
I don't think it's super surprising. People from different regions who speak the same language use some words and phrases in different frequencies, like the text equivalent of mild accents. And that's exactly the kind of thing it'd be easy for a LLM, trained on word frequencies from a ton of text, to pick up on. And then just make up the 'reasons'.
The thing is, I tried it on several other LLMs. O1 and Opus declined to answer on ethical grounds. Gemini 1206 and Flash 2 failed. Smaller OSS models failed too.
I think it's fair to say that Claude is uniquely good at this. I'd wager superhuman at the task.
The reasons it gave aren't anything I would have picked up on myself barring an Indian streak towards pragmatism, but I remember Dase making similar sweeping observations in the past which modestly boosts my confidence that Claude is being honest in its self-reporting.
In the initial attempt, correctly identifying that an immigrant doctor to the UK is most likely to be South Asian is a good catch, and the other models faltered. This was removed and other comments substituted that only identified I was a doctor, which made it lean more towards me being American, but still of Indian origin.
So I expected LLMs to be "okay", somewhat better than I could. It turns out that that's not the case here, and Claude beats any human who isn't Rainbolt cracked and other LLMs.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link