Given that gaelan exercised much control over greg by registering him amd by curating his training data, i think gregs actions should be considered just an extention of gaelan
On Wed., 22 Jul. 2020, 10:15 am Gaelan Steele via agora-discussion, < agora-discussion@agoranomic.org> wrote: > Alright, people seem to have started to get annoyed, so I believe > continuing this experiment would violate my sacred and eternal duty to > Treat Agora Right Good. > > So, yeah. I'm Greg. Or maybe I'm sending messages on Greg's behalf. See > recent CFJs. > > As most of you accurately surmised, Greg's messages were generated by > GPT-2, specifically a version of GPT-2 fine-tuned with Agoran mailing list > logs since 2014. (The 2014 date is largely arbitrary—I might have been able > to go a bit further back, but using much more data ran into resource > limitations.) > > Greg was implemented using a combination of shell scripts, commands in my > shell history, python scripts, a Google Colaboaratory Notebook, and me > manually copy/pasting messages around. Notably, I manually pasted in and > hit send on each message. I did this primarily because figuring out email > APIs sounded like a PITA, but also because I wanted to be able to pull the > plug in case it said anything horrific. I did, however, do as much as I > could to avoid injecting my free will into the process. I operated off of > two rules: for each message to the public forum, I would run a python > script which had a 10% chance of invoking GPT-2 to generate a reply, which > I would send verbatim. (GPT-2 barfs on overly-large input data, so I > included a failsafe that automatically removed old messages in the thread > until the input was small enough to work. Some messages (like the rulesets) > were far too big on their own, resulting in the code generating "replies" > without any context.) Additionally, each day after the first (which looks > like is just going to mean "today"), I ran a script which had a 50% chance > of generating a brand-new proposal. Had I been aware of CFJ 3790, I might > have actually went to the trouble of having it send the messages > automatically after generating them. > > I did "intervene" twice: for the registration message, I specifically > asked GPT-2 to generate a message to BUS with a subject of "BUS: > Registration". In my testing, this had about a 75% chance of generating a > message that was a somewhat plausible attempt to register. Unfortunately, I > got unlucky and my first "real" attempt to generate a registration message > resulted in something completely random (a proposal, I think), so I > generated a second one and sent that one. In another case, I discovered a > bug with the large-input failsafe (turns out, GPT-2 can barf by silently > returning the input with no additional output, or by throwing an exception; > I was only handling the first case), so I fixed the bug and re-ran the > generation. In every other case, I mechanically copies messages back and > forth, following the plans I had made before sending the first message, > without attempting to impose any editorial control. > > In my testing, Greg did occasionally borrow other people's signatures, but > I didn't expect it to be this common. I considered preventing it from doing > this by removing signatures from the training data (so it would never learn > to include them), but I thought it was rare enough and amusing enough that > it wasn't worth removing. In retrospect, I probably should have removed > them. > > In my testing, I ran into several outputs that were interesting enough to > save so I could show them later. Here are links to them: > > A proposal for something vaguely resembling a functional auction mechanic: > https://gist.github.com/Gaelan/e7f7d3fc48c1abd08f0afb8049077acb > A made-up FLR excerpt containing an interesting-sounding royalty mechanic: > https://gist.github.com/Gaelan/8d092a17ed9c210685a4f4dd1e622ae2 > Another ruleset excerpt, containing the core rules of an alternate > universe Agora: > https://gist.github.com/Gaelan/ee631f9f97b53df8483e342ef36b6618 > A batch of attempts at starting new threads, with varying quality: > https://gist.github.com/Gaelan/0c027e3f5b97dab700182aa663401f47 > A fake rule called the "Register of Proposals", which looks like a > semi-plausible implementation of proposals in an alternate-universe Agora: > https://gist.github.com/Gaelan/0c6853f500799c5190a0a1ef474b098b > > I'd be happy to share the model at some point, but its a bit of a pain—I > think it's 1.5 GB—so I'm not sure how best to do that. In the meantime, I'm > happy to try out any inputs y'all are curious about. Also, From headers > were included in the training data, so I should be able to ask it to > generate message from specific Agorans. That might be fun. > > Happy to answer any questions, of course. > > Gaelan > "Nothing in a democracy is sacred, and nothing in a democracy is > sacrosanct. That's not to say it's never been broken, but it's been > tracked for quite some time." > [GPT-2 put that quote in someone's signature during one of my tests. As > far as I can tell, it's not a real quote, but it sounds like a fairly > interesting reflection on nomic.]
