Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.
- 171
- 1
Jump in the discussion.
No email address required.
Notes -
Claude Opus 4.7 knows who I am, by name, and without access to web search.
It also pegged me more often than not from an excerpt of text I'd written half a decade back, once again, without internet access. Well, fuck. I did always harbor aspirations of becoming famous enough a writer to be known to LLMs by name, but this also confirms my previously stated belief that privacy on the internet is on the way out. Pseudonyms won't save you, stylometry is all you need.
I was startled a month or so ago when I pasted in a thread from the Motte as context to ask about sports ticket policy (from after its training cutoff), and in its answer, it casually used female pronouns for @2rafa, who had made a short comment that was not at all revealing of gender.
How tf does the model know that? No one in that thread had referred to her gender. I triple-checked. It must be in the weights somewhere. It must have picked it up over the years of people using "she/her" in reference to that handle, and just follows suit.
I wonder if it has this kind of built-in knowledge of all minor internet hangouts, or is the Motte special? - elevated as a high quality source explicitly or implicitly
The most likely answer is surely just that all our writing is in the training set. Most of the internet is, including all of Reddit.
Certainly there are frontier lab researchers here, a few lurkers. Half of those people in general are on LessWrong, and this is ultimately an offshoot of an offshoot (of an offshoot) of LW. But is that the reason? I doubt it.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link