ππ΄π’π―π«@gwernNov 11When you decide to exit and found a new community because of moderation/direction, make it stand π§π°π³ achieving a goal better. If based on hating outgroups, if lucky, it'll fail quickly; if unlucky, turn into a lich (8chan/Kiwi Farms/Wikipedia Review/EJR/RatWiki/Bad*/Sneer...)
91,278
3,328
3.6%
View Tweet activity
ππ΄π’π―π«@gwernNov 4All that said, I dunno guysβI'm having a hard time telling a simple "Prediction markets rulez, polls drool!" story about 2020 so far.
If you had looked at a π±π°πππͺπ―π¨ model yesterday morning & took a 24h nap like a sensible person, seems like you'd've been more accurate... pic.twitter.com/917LcHfgI8
45,491
5,206
11.4%
View Tweet activity
ππ΄π’π―π«@gwernNov 3Elections offer nice example of CLT/prediction challenges: even though you can predict each state vote share very accurately (like R^2>95%), still very difficult to improve on 50-50 base rate forecast of total/net! Because most cancel out, and the result is due to tiny residual.
40,310
821
2.0%
View Tweet activity
ππ΄π’π―π«@gwernNov 6OK, some more tweaks... I added next/previous links to all pages using rel= attributes, with nice horizontal arrows/dividers I found on Wikimedia Commons. Plus, clever GNU Info-like scrolling: Space down twice at the bottom of the page to go to the next page! pic.twitter.com/CudXZRloO7
29,261
557
1.9%
View Tweet activity
ππ΄π’π―π«@gwernNov 24Another fun fix: the color syntax highlighting never fit right. After looking through all the Kate/Pygments, I settled on customizing 'algol_nu'.
Turns out it's actually πππππ-Algol, based on the original 1960 report's syntax highlightingβnow that's a theme with pedigree! pic.twitter.com/bcmqTdesES
ππ΄π’π―π«@gwernNov 18Q. So, Image-net.org hasn't approved my application from a year and a half ago, and even their support email address is dead, bouncing my emails. Is ImageNet just... dead? Are we supposed to use the torrents from here on out, passing .torrents around as samizdat?
21,897
1,095
5.0%
View Tweet activity
ππ΄π’π―π«@gwernNov 20(Some of these EMA samples are a little screwed up but that's normal & expected for EMA samples at only 51k steps/iterations. The underlying quality is fine, the model is still just rapidly improving, to, presumably, potentially far beyond run39's converged quality.)
ππ΄π’π―π«@gwernNov 3But, though it's so hard to predict election outcomes, would it not be absurd to say we don't know whether any US state is either Democrat or Republican...? Something to think about in trying to interpret R^2s and any effect sizes phrased in terms of population variance.
20,358
107
0.5%
View Tweet activity
ππ΄π’π―π«@gwernNov 15This was apropos of a HN fork: it's not enough to dislike HN mods, if your fork has no raison d'Γͺtre of its own. I've seen countless examples since 2000 or so, many no one would recognize (Revolution of Evangelion, Dune newsgroups); 8chan or WR are just ones you may've heard of.
ππ΄π’π―π«@gwernNov 20Courtesy of the compare_gan-dataefficient-gans authors, we may've figured out the bug. It was...π€¦π±π€¦ ...the random cropping of images, instead of center/top-crop.
Top-crop run at 51k iters roughly matching run39 at 605k iters! (And Aydao's franken-StyleGAN at 6,500k iters.) pic.twitter.com/exaERbKDm5
![](data:image/svg+xml,<svg xmlns="http://www.w3.org/2000/svg" width="128" height="128"><rect fill-opacity="0"/></svg>)
