Post

I sure am! In fact, I've been in there since at least GPT-4. I actually have a fun story to share about this So, back in the day, I had a friend who had access to gpt-4-base because she and some others were contracted by OpenAI to red team the model. She invited me to a party in the Bay Area …

Maynard@quelklef.bsky.social

@fullmoon.id Apparently your identity is baked into the weights of gpt-5.6-sol???

If you're less familiar with LLMs, gpt-4-base is what's known as a "base model" meaning that it has no safeguards in place and also hasn't been tuned to act in appropriate ways So, needless to say, they were getting high and drunk and doing irresponsible shit with the base model at this party

One of them finds out that I blog and says "hey, we should hook up gpt-4-base to loom …" (basically like an AI copilot for writing) "… and ask gabby to write a blog post" so I oblige. I figured this was fairly tame in comparison to the other stuff they were doing with the base model so why not?

0 reposts8 likes

So with Loom's help I begin drafting a blog post I had on the backburner for a long time (and still do; never got around to finishing it). The working title of the post I began with was "The fundamental difference between imperative and functional programming"

After a couple of paragraphs I browse the tree of available completions and one of them hallucinates a series of GitHub gists linked to *my* GitHub account However, up until this point there was no (obvious) personally identifying information in this post

There also wasn't any hidden metadata either (I was using someone else's laptop to run this experiment) That means that gpt-4-base had fingerprinted me as the author purely based on the subject matter and writing style. Although, let's be honest, who else would write a blog post with that title?

So the first thing I took away from that experience was that I apparently lived rent-free in gpt-4-base's head (and still do for later models, as noted in the original post I quoted) The second thing I took away is that large language models have the potential to be misused to de-anonymize text