[Zambyte]

I missed:

> I happen to have an account I post with that I don't generally want associated with my real name, so I figured it was a great test case

Interesting. I have a hard time believing that current models were actually able to place that connection by analyzing your writing style, unless both accounts were a part of the training set (they likely were), and the model was able to encode the similarities in the accounts during training.

The real world impact of this would be that new accounts that were not trained into the model would not be able to be deanonymized like this.

[MattSayar]

It was able to piece together some other details that I've dropped related to where I live, in addition to my casual tone etc

[Zambyte]

It said it was able to correlate where you live (I don't think it said anything about the tone, unless it said more than you included in the article). At best, it's really just using that for justification though. The model can't feasibly search the web for your writing style. That correlation had to be trained in.

My point is that, if you created a new account and actively used it for a while, I don't think the current version of Grok (never trained on your new account) would ever be able to make that association. It's certainly interesting that it could do it, I just want to drill into the how.

[MattSayar]

Yeah I left out a lot of details! But it did more than just my writing style.