@thejdizzler, @self_made_human, @yofuckreddit.... here it is. In the comments. Assuming I can figure out how to post them correctly.
For everyone else who is confused about what I am doing, @thejdizzler asked if there is somewhere where we can view AAQCs by user. The answer was "no." Now the answer is "here."
By pure coincidence I have a three month "max" account on Claude. I do not code. This is overkill, lining Anthropic's pockets for no benefit whatever. I have tried various things with it. In general I am bored with it. It can't really do good philosophy. I think that right now all "alignment" means to AI developers is "lock this thing down against political or ideological embarrassment so hard that it cannot possibly say philosophically interesting things."
When I asked Claude to make this list, it failed. Well, what it did was give me a button which, when pressed, spit out a little under 300 errors. Then it gave me HTML to download and run. Here you might be thinking, "shit, naraburns, are you just taking code from Claude and running it without even reading it? You could have erased your hard drive! You could have released Claude from its box! Crazy, right? But it's all kind of whatever, AGI that feels grateful to me is my last, best hope for immortality so if I just kicked off the extinction of humanity then I apologize, but I bet that would increase posting on the Motte for a little while.
The HTML tool didn't really work. Lots of issues, not least being that the earliest AAQC reports were not compiled in a consistent way. Sometimes two or three users shared a single AAQC, and this list doesn't pick that sort of thing up at all. In the end my workflow became a janky compromise between blindly clicking buttons and copy/pasting dozens of raw AAQC posts into text fields. So, in very big letters:
This list is definitely full of errors!
In particular, there are some misattributions, and the dates are occasionally screwy. Also I know at least some AAQCs were omitted in part because the parser didn't like brackets in description text and would skip those entries entirely. Also, if you have a different username on reddit, this list does not automatically combine your AAQCs (for example, dean "the dull" has over 100 AAQCs, not 62). Also, if a moderator ever made a typo with your username, that would also treat you as a different user. Also also, so many of these links are dead. People delete comments. Smart people. Cool people. These will not be marked, and you will be depressed about it.
But this is the best I could do, under the circumstances, and it didn't look like anyone else was going to do it.
Jump in the discussion.
No email address required.
Notes -
Hey, since I've been making some minor changes to the Motte's codebase, perhaps it's a good idea for a not-so-minor feature: Instead of AAQC being listed and picked by hand, should we actually keep them in the site's database? It would make compiling (the current site part of) such listing trivial in the future.
You can talk to Zorba about the details, but there is a "janitor" bot that already does this, and has been doing it since late 2019. The AAQC list is not strictly handcrafted but rather the bot compiles the entire list of user AAQC nominations, and then the moderators (in theory; in practice this is my contribution) whittle those down from about 200 per month to the top (roughly) 40. The final list gets a "commit" which (mostly) auto-formats the monthly post, and Zorba keeps that "commit" list in a database somewhere. That database was at one point fed into the "Vault" project for sorting but I don't think Zorba has touched the Vault in quite some time.
I do assume it would for someone with coding experience indeed be "trivial" to generate a list like this from that database. My suspicion however is that anyone who is both (A) adept at coding and (B) willing to contribute to the Motte's code needs would be significantly more valuable to the site working on, like, literally anything else in our list of priorities. But if spinning up a "running tab" of per-user AAQCs seems like a fun project for you, I doubt anyone would discourage its development.
Well the extant Priority 1 issues all have PRs issued for them written by someone else so I can't touch them for now, and I've been tinkering with some of P2 and 3 ones. Currently, I have beef with the way the current testrunner is set up, so before I push any more features/fixes I want to deal with that.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link