I believe that TSP had this post prepared (at least the claim prepared) long prior to me ever posting in his meta thread about QS. He wanted an excuse to have run the "tests" he run in the first post. His conclusions are far from sound considering the fact that his overall sample was very small. I wonder how many tests TSP had to run until he found three others who have very high numbers against eachother
Quickseller, you admitted in the third post of this thread that you didn't really understand the experiment. That's okay, it is a little complicated. Thankfully for you, you can download all the code and run as many tests as you'd like.
This is the setup:
1) I saw that Quickseller had written many, many posts under his main account, the corpus I downloaded was approximately .5 million words after tokenization
2) I knew that Quickseller had written a substantial amount of posts under an alt account which he stopped denying was his a few months ago (ACCTSeller).
3) I had more than a strong intuition that Panthers52 was QS' account, and he's written a prety substantial amount of posts with that account too.
4) The objective, then, it to quantify the similarity of the language of the Panthers52 account with repect to the known alts.
So, the experimental hypothesis is a two parter:
A language model trained on the posts of Quickseller will:
* predict the corpora of ACCTSeller and Panthers52 with equivalent accuracy
* predict the copora of ACCTSeller and Panthers52 with much better accuracy than the corpora of people who aren't his alt
The results:
The experimental results support both hypotheses quite strongly. Even the model trained on the ACCTSeller data (which is considerably less robust than the QS one because of the smaller size of training data) predicts the QS and Panthers corpora much better than it predicts the copora of the non-alts in the study.
Because the code for the experiment has been made public, anyone is free to replicate the results or to run their own experiments. If QS really wants me to run his model against more and more accounts, it's my intuition that he'll just be putting more nails into his own alt's coffin. Maybe I'l do it tomorrow, I don't really feel like doing it today. What's more, the behavoir of Panthers52/Quickseller both in this thread and also
here already shows that they're alts to anyone who takes time to look into it. So, to me, the quantitative data is sorta just the icing on the cake, so to speak.
There are actually a few other metrics used in textual studies which I think would be fun to apply. Stuff with respect to hapax legomena. If I have time, I'l try to brush up on some of these modelling techniques and I can release some further quantitative measures. But, alas, I actually have some others stuff to do this weekend so it may be a few days.
Anyone who wants to know more about the experiment and/or receive the language models generated should PM me or email me.
The question for those on default trust at this point is this:
Given how quick you guys are to mark red trust on anyone even suspected of scamming and how it's basically considered around here that the burden of proof is on the accused to prove themselves innocent, isn't it time to warn others about trading with Panthers52? That, at the very least, someone trading with Panthers52 might be well advised to use an escrow other than Quickseller if he really wants a neutral third-party escrow?