|
pointless disagreement that will cause a split in the movement is an accurate representation of the left.
|
#
?
Oct 24, 2016 02:40
|
|
|
|
| # ? May 15, 2024 08:51 |
|
|
https://twitter.com/WYR_bot/status/790402464050737154
|
|
#
?
Oct 24, 2016 05:22
|
|
|
https://twitter.com/WYR_bot/status/790221618849144832
|
|
#
?
Oct 24, 2016 08:11
|
|
|
fermun posted:pointless disagreement that will cause a split in the movement is an accurate representation of the left. lol 
|
|
#
?
Oct 24, 2016 10:02
|
|
|
hey crusader have you considered another nn that trains on wyrbot's questions and responses? i.e., you feed it the questions and have it predict the winner (or even the percentage split) then compare with the actual results. i don't know if there's enough training data but it would be interesting to see if you can become at all accurate at guessing which answer will win in advance. assuming it worked you could even use it to adjust wyrbot's questions, for instance if you wanted to try and ask fewer one sided questions. could even do something similar with the number of votes each question gets to try and ask better/more popular questions even if you don't want to alter wyrbot (why mess with perfection) it'd be cool to see if a network can "understand" human preferences on weird wyr questions
|
|
#
?
Oct 24, 2016 13:44
|
|
|
yeah, it's a reinforcement learning problem i think (e.g. use some delta threshold as a win condition), but i haven't attempted it since i figured i'd need thousands more results - and, to your later point, i also haven't been sure about close results necessarily being what i want to skew towards because sometimes the lopsided ones are funny... will definitely keep exploring to see what's possible though
|
|
#
?
Oct 24, 2016 13:52
|
|
|
yeah, the obvious ones are great and i don't understand why you'd want to get rid of them
|
|
#
?
Oct 24, 2016 14:41
|
|
|
teaching a bot to predict the winner seems like a fun thing to do though, use it on other polls or something idk btw are there other good mashups like erowid recruiter? https://twitter.com/erowidrecruiter/status/560559080289222656 https://twitter.com/erowidrecruiter/status/767769577895768064 https://twitter.com/erowidrecruiter/status/759914312408240129 https://twitter.com/erowidrecruiter/status/759189274709233665
|
|
#
?
Oct 24, 2016 15:09
|
|
|
my girlfriend is Legos posted:yeah, the obvious ones are great and i don't understand why you'd want to get rid of them it was just an example, i'm not sure that there's actually a percentage breakdown that you'd want to target
|
|
#
?
Oct 24, 2016 15:36
|
|
|
Snapchat A Titty posted:teaching a bot to predict the winner seems like a fun thing to do though, use it on other polls or something idk This bot is my actual life.
|
|
#
?
Oct 24, 2016 15:56
|
|
|
are there neural network architectures besides LSTMs/RNNs that can do generative text? like GANs or something
|
|
#
?
Oct 24, 2016 23:34
|
|
|
you could for sure use a GAN approach but you could probably even use a CNN - check out deepmind's PixelCNN and WaveNet architectures, seems like something analogous would work for text
|
|
#
?
Oct 25, 2016 00:02
|
|
|
train it based on # responses instead of polling scores and then gin up an adversarial network to manufactuer the most viral memes of all time
|
|
#
?
Oct 25, 2016 00:08
|
|
|
marketing companies are probably even now hiring ML researchers to design networks that generate the most optimally dank and viral memes of all time rumour has it that DennysNet has already become self aware
|
|
#
?
Oct 25, 2016 00:13
|
|
|
loss.jpg function
|
|
#
?
Oct 25, 2016 00:14
|
|
|
https://twitter.com/WYR_bot/status/791308565701074944
|
|
#
?
Oct 26, 2016 17:29
|
|
|
https://twitter.com/WYR_bot/status/790764909516689408 i like this one
|
|
#
?
Oct 26, 2016 17:39
|
|
|
lmao
|
|
#
?
Oct 26, 2016 17:42
|
|
|
wyrbot is eerily good
|
|
#
?
Oct 26, 2016 17:43
|
|
|
Snapchat A Titty posted:wyrbot is eerily good tbh if I found out the wyrs were generated but then human-curated to pick the best I wouldn't be mad.
|
|
#
?
Oct 26, 2016 19:15
|
|
|
NoneMoreNegative posted:tbh if I found out the wyrs were generated but then human-curated to pick the best I wouldn't be mad. oh, its absolutely curated, but the curation is for heinously offensive ones (there was reddit in the corpus for variety but that sometimes yields lines i'd rather not make it to twitter), or ones that just make no sense at all - the net still doesnt have a firm handle on longer sentences unfortunately. beyond that i dont want to try and predict which ones people would enjoy myself - i've already seen many that people enjoyed that i had thought were just doofy (will not be surprised at all if a future version is better at predicting laughs than me). these days i generate a month's worth from the rnn, scan it for crud, then pass it on to the twitter script - the monthly cadence works well with re-training on additional corpus material. eventually with more data and real thorough scrubbing of the corpus it maybe could just post on its own; with lower temperature (higher confidence) settings right now the wyr's are grammatically correct but gravitate towards the same few topics (flying, reading minds, pets) - i want it to be confident -and- creative. using the # of responses as feedback is also cool; will need to ascertain how to weight against the # of followers and/or impressions to give earlier ones a fair chance - i wonder if I should just re-post ones from prior to the polling also. anyway, as always, suggestions welcome!
|
|
#
?
Oct 26, 2016 23:02
|
|
|
"heinously offensive ones (there was reddit in the corpus for variety but that sometimes yields lines i'd rather not make it to twitter)" this ain't twitter, mang - make w/ the offence 
|
|
#
?
Oct 26, 2016 23:04
|
|
|
Crusader posted:oh, its absolutely curated, but the curation is for heinously offensive ones (there was reddit in the corpus for variety but that sometimes yields lines i'd rather not make it to twitter), or ones that just make no sense at all - the net still doesnt have a firm handle on longer sentences unfortunately. beyond that i dont want to try and predict which ones people would enjoy myself - i've already seen many that people enjoyed that i had thought were just doofy (will not be surprised at all if a future version is better at predicting laughs than me). these days i generate a month's worth from the rnn, scan it for crud, then pass it on to the twitter script - the monthly cadence works well with re-training on additional corpus material. Was the original corpus just text, or a wyr-type subreddit?
|
|
#
?
Oct 26, 2016 23:12
|
|
|
NoneMoreNegative posted:"heinously offensive ones (there was reddit in the corpus for variety but that sometimes yields lines i'd rather not make it to twitter)"
|
|
#
?
Oct 26, 2016 23:15
|
|
|
NoneMoreNegative posted:"heinously offensive ones (there was reddit in the corpus for variety but that sometimes yields lines i'd rather not make it to twitter)" deeeeeerp, yeah, i should have been at least posting them here all along  im traveling this week without the aws ssh key but i'll see what i can find from past runs or will just start keeping them on future runs.
|
|
#
?
Oct 26, 2016 23:16
|
|
|
atomicthumbs posted:Was the original corpus just text, or a wyr-type subreddit? it's all normalized text ("Would you rather " as the exact starting string, lower-case all additional text, don't really care beyond that as long as there's an " or " somewhere) from everywhere i could find wyr posts in bulk - the reddit part are the topics from the wyr subreddit, but there's buzzfeed, pdfs of "would you rather" books, and i forget all what else in there cinci zoo sniper posted:ehhhhh, no or not, either is fine with me
|
|
#
?
Oct 26, 2016 23:20
|
|
|
i mean i'm fine with just offensive stuff, before you mentioned source there was no telling if it doesn't construct posts from arbitrary reddit lot where you can run in a whole load of awful garbage beyond "would you rather browse in amberpos or becomes allergic to beer"
|
|
#
?
Oct 26, 2016 23:26
|
|
|
do we have a list of the yosbots? mine is not set up properly bc of eponymous reason: @laziestboy
|
|
#
?
Oct 26, 2016 23:34
|
|
|
fine work, forums user Crusader
|
|
#
?
Oct 26, 2016 23:35
|
|
|
Crusader posted:deeeeeerp, yeah, i should have been at least posting them here all along you could always tweet out monthly pastebins/tweetlongers/whatever of the outtakes without putting them in the regular rotation
|
|
#
?
Oct 26, 2016 23:39
|
|
|
fuckin hell https://twitter.com/WYR_bot/status/791127348649263107
|
|
#
?
Oct 30, 2016 08:54
|
|
|
Where theres a will theres a whey
|
|
#
?
Oct 30, 2016 11:02
|
|
|
https://twitter.com/WYR_bot/status/792667709486067712 https://twitter.com/WYR_bot/status/792486493705084928 (summer) thank you wyrbot unpacked robinhood fucked around with this message at 11:35 on Oct 30, 2016 |
|
#
?
Oct 30, 2016 11:31
|
|
|
https://twitter.com/WYR_bot/status/782596036900614144 https://twitter.com/WYR_bot/status/792305272165429248 https://twitter.com/WYR_bot/status/793573798104670208 https://twitter.com/WYR_bot/status/761577723324227584 wyr bot why do you want me to eat butter
|
|
#
?
Nov 1, 2016 23:31
|
|
|
https://twitter.com/AdobeCare/status/795560202405867520?cn=cmVwbHk%3D
|
|
#
?
Nov 7, 2016 21:29
|
|
|
If the fine tooth comb they're offering for the bot's lice is anything like the one they use to find other awful bugs, well...
|
|
#
?
Nov 7, 2016 21:43
|
|
|
https://twitter.com/markov_polov/status/795548365127749632 https://twitter.com/markov_polov/status/795550896268537856 https://twitter.com/markov_polov/status/795724626869428224
|
|
#
?
Nov 7, 2016 22:57
|
|
|
|
|
#
?
Nov 7, 2016 23:27
|
|
|
|
| # ? May 15, 2024 08:51 |
|
|
It hurts, Trig… …I’m h…a…p…p…y… …friends… …It hurts, …it hurts… Trig, Trig, Trig, Trig, Trig, Trig, Trig, Trig, Trig, Trig, Trig, Trig, Trig, Trig, Trig, Trig, Trig, Trig, Trig, Trig, Trig, Trig, Trig, Trig….. …go… b… a… c…k… …Trig… I feel… g… o… o… d… …I’m so sad… …..Trig Trig! Ah, Grrr, Ohhh… Argh… Yaaagh!… It’s not right… not right… not right…
|
|
#
?
Nov 7, 2016 23:37
|
|