ππ΄π’π―π«@gwernApr 9Another day another WP entry 'improved' to remove interesting sourced text I refed (disappearing polymorphs here): en.wikipedia.org/w/index.php?tiβ¦
Do WP articles still improve, or do they just gradually rot away these days? Any longitudinal research? Maybe I should just link revisions.
25,817
668
2.6%
View Tweet activity
ππ΄π’π―π«@gwernApr 2I always enjoy the CS/ML/math/statistics satire in the annual SIGBOVIK; physics has Arxiv April Fools papers & journals like Special Topics; and medicine has the BMJ Christmas issue, of course.
What are the equivalents in the humanities, like sociology or literature?
ππ΄π’π―π«@gwernApr 10Someday I should take 20 random WP articles I link, and do a 5-years-ago/now comparison and see if they improve. If they don't on average, might as well link to the revision...
Also going to start logging anchor-not-found errors - deletionists love renaming/deleting sections.
ππ΄π’π―π«@gwernApr 14Less pleased by archiving code revisit: 200+ links didn't archive & need manual reviewing+fixing.
Unpleasant Internet archiving truth: as abusive, neglectful & malicious as Big Tech websites can be, still longer-lived than academia/hobbyists/govt etc.
(That saidβfuck Medium.)
13,387
88
0.7%
View Tweet activity
ππ΄π’π―π«@gwernApr 17Whatever happened to super-convergence? arxiv.org/abs/1708.07120 LRs 3β20 sometimes save >50% of compute. No one seems to know when or why it works. ~No followup despite Fast.ai pushing it, aside from occasional successes.
Has anyone tried on ππ’π³π¨π¦ models?
12,744
425
3.3%
View Tweet activity
ππ΄π’π―π«@gwernApr 13I've also noticed that Twitter search has become almost entirely useless. It only goes back a few days now (β½), and fails to pick up all sorts of hits I π¬π―π°πΈ exist (eg try gwern.net URL searchs with 'url:' etc; I can see plenty in analytics, and yet...) twitter.com/AlvaroDeMenardβ¦
ππ΄π’π―π«@gwernApr 24(Incidentally, not a single person even tried to rise to my challenge to name 3 Chinese DL research results from the past few years on par with just OpenAI's output over the past few months like CLIP/DALL-E.
No, sorry, 'rotary embedding' doesn't count.)
ππ΄π’π―π«@gwernApr 17Given how linear and stable training curves for models like CLIP & GPT-3 are, and the tunneling interpretation of super-convergence for highly overparameterized models (old.reddit.com/r/MachineLearnβ¦), seems like it might work for super-large models. At possible <90% compute savings...
8,450
124
1.5%
View Tweet activity
ππ΄π’π―π«@gwernApr 10TIL apt will happily add libc6 to its removal list in an apt-get dist-upgrade, thereby not just committing suicide when it crashes but also making every executable on the system not runnable. π
On the plus side, now I know how to mount software RAID & unpack .debs into a mount.
8,120
64
0.8%
View Tweet activity
ππ΄π’π―π«@gwernApr 1Initial notes on how I think collecting/collectibles works (not the high-end, which I see as a mostly different and only superficially similar phenomenon, but the stuff ordinary people do): gwern.net/Collecting
ππ΄π’π―π«@gwernApr 15I regularly get complaints on desktop about the font size being too small / text hard to read, for body & popups, so we've boosted the base text size from 16 to 20. (Mobile apparently is fine.)
Guess we'll see if people start complaining about it all being *too* big now.
![](data:image/svg+xml,<svg xmlns="http://www.w3.org/2000/svg" width="128" height="128"><rect fill-opacity="0"/></svg>)
