𝔊𝔴𝔢𝔯𝔫@gwernJan 29So by my count, just over the past month, OpenAI has seen a diaspora of: Dario Amodei, Sam McCandlish, Tom Brown, Tom Henighan, Chris Olah, Jack Clark, Ben Mann, and Paul Christiano (just the publicly known ones).
OpenAI, are—are you alright? Do you need to talk to someone?
36,027
5,370
14.9%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernDec 31The precautionary vs proactionary principles: "He that will not apply new remedies must expect new evils; for time is the greatest innovator."
𝔊𝔴𝔢𝔯𝔫@gwernJan 11If you want to feel old (intellectually), consider this:
Someone who started their PhD because they were excited by the publication of this new 'ResNet' thing would be just finishing, given 5-year averages.
(And there are people finishing recently who started around word2vec.)
𝔊𝔴𝔢𝔯𝔫@gwernJan 26Added tag links & metadata blocks.
I think you can reach any page on gwern.net through popups, if you are careful, but you 𝘱𝘳𝘰𝘣𝘢𝘣𝘭𝘺 can't popup 𝘦𝘷𝘦𝘳𝘺 page simultaneously because of the deduplication logic & dead-ends like the directory index pages... pic.twitter.com/mZVOF06612
20,832
252
1.2%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernJan 31Hacktivism idea: high leverage: write very polished smartphone book scanning app, which can download metadata+call-no of books in local library missing from LibGen, assign you 3, & upload w/metadata. Coordinate global book-scanning flashmobs. (There's not *that* many books...)
19,930
990
5.0%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernJan 11(Because it's so expensive in multiple ways, a CNN, which could be run on every possible combination to provide kerns if desired, may well produce better results than standard hand-kerning of a few thousand pairs!)
𝔊𝔴𝔢𝔯𝔫@gwernJan 13Today in 'really, just try to google it':
x> "anybody here know that one artist with the really creepy art sytle that starts with a z?"
I googled: 'that one artist with the really creepy art sytle that starts with a z'.
Hit #2: en.wikipedia.org/wiki/Zdzis%C5%…
𝔊𝔴𝔢𝔯𝔫@gwernJan 31All of the necessary software & machine learning tools for getting high-quality book scans from smartphone cameras has existed, AFAICT, for a decade or so. But no one has broken the bottleneck of putting it all together. Why can't we have millions of people casually scan a book?
13,942
178
1.3%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernJan 17We removed the auto-scrolling. Apparently, no one has the slightest idea what I mean by 'GNU Info-like scrolling'. (Darn computing pop cultures!)
13,446
87
0.6%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernJan 11Idea: CNN for font kerning. By far most tedious, labor-intensive/time-money-expensive, rote part of high-quality font creation; but also most I-know-it-when-I-see-it, non-semantic, and easy to create n>millions by just jittering random text set in existing kerned fonts (k>50k).
13,408
89
0.7%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernJan 31(Thought triggered in part by memory of Larry Page pointing out the feasibility of Google Books by timing himself flipping the pages of a book, and assuming software can deskew a good camera's video. Well, guess what we all have in our pockets today, thanks in part to Page...?)
12,763
140
1.1%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernJan 12Site usage poll:
on link popups, there are helper links to the Internet Archive and Google/Google Scholar, in case the reader wants to go to an IA archive or do reverse citation links (either by DOI or by `link:` in Google).
Have you ever clicked on one & found it useful?
12,666
80
0.6%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernJan 14They pooh-poohed transfer learning or RL or one-model-to-rule-them-all. All near-term AI would be like Deep Blue: superhuman but exquisitely-specialized. We'd have thousands of Tool AIs each doing tasks like doctoring, but never any kind of 𝘨𝘦𝘯𝘦𝘳𝘢𝘭 model.
12,177
173
1.4%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernJan 12We added logging yesterday, and so far there appear to have been 0 clicks, suggesting <1/8000 per popup usage rate. The JS+CSS code is kinda annoying so we're seriously considering removing it if literally <1 person per day even uses them... pic.twitter.com/BEKiOtzNz1
11,121
271
2.4%
View Tweet activity
Engagements
Showing 31 days with daily frequency
Engagement rate
2.4%
Jan 31
3.3% engagement rate
Link clicks
4.3K
Jan 31
77 link clicks
On average, you earned 139 link clicks per day
Retweets without comments
27
Jan 31
0 Retweets without comments
On average, you earned 1 Retweets without comments per day