𝔊𝔴𝔢𝔯𝔫@gwernDec 21Trippy to think about how "a large part of the earth is still in the Urne unto us"; we know for centuries there's a library buried under ash & lava containing more Greco-Roman works than the entire surviving corpus. In a century we'll read it, and half of classics will be toast.
𝔊𝔴𝔢𝔯𝔫@gwernDec 12I have also learned the hard way that the API docs aren't kidding about avoiding newlines. Bizarre: I added (semantically relevant!) newlines, and my ada/curie results became practically random and had to rebuild them completely.
Why do newlines screw up the embedding so much?
8,592
106
1.2%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernDec 26Did you think I couldn't find another way to add more hyperlinks? 𝙒𝙍𝙊𝙉𝙂❗
I realized I could hack up the existing similar-links code & Emacs formatting function to append them automatically when writing a new annotation, to be manually approved. pic.twitter.com/m0MEQENcv1
𝔊𝔴𝔢𝔯𝔫@gwernDec 22twitter.com/gwern/status/1…
(All we know of the present is that it is already hopelessly quaint, just obsolete things & ignorant ideas perched over the infinite abyss of reality, to be forgotten, and that way of life forever gone.)
6,515
150
2.3%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernDec 14Gratz to Elon Musk for winning 𝘛𝘪𝘮𝘦 Person of the Year a second time; but he's got a ways to go to match Barack Obama's 3–5 𝘛𝘪𝘮𝘦 awards.
6,425
141
2.2%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernDec 18Wondering why syncing slow, I realized that my Pandoc executables were 3 orders of magnitude larger, each, than the hard drive of my first computer. Not everything is exponential but boy howdy can it matter when they are.
Exponentials are a helluva drug.
twitter.com/chrischirp/sta…
6,018
70
1.2%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernDec 6Current Status: going "oh no oh no oh no" that I can't just write out gigabytes of doubles as strings and parse them in Haskell without any issues.
5,857
69
1.2%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernDec 17Consider this a tip for any data detectives out there who needs a post-ivermectin hobby. I'd start with Satish's co-authors and ask about what they saw firsthand, exactly, and where the data comes from in all these papers.
𝔊𝔴𝔢𝔯𝔫@gwernDec 25UK Biobank immediately comes to mind as underrated, even among the human geneticists for whom it is the sun they bask in and air they breathe.
5,010
96
1.9%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernDec 24Keep going! You might get 'grokking' ! Winners never quit, and quitters never win.
4,771
50
1.0%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernDec 21It's strange to read a paper about, eg, Julius Caesar's last words or about the Antithykera mechanism or Barbegal or the Nemi ships, and have to append a mental asterisk: "...But of course, all of this is wildly ignorant guesswork, and will probably be trashed post-Herculaneum."
4,679
82
1.8%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernDec 21(You see the number 1,800 scrolls, but that's just the number retrieved from just the one alcove of Herculaneum! Think of how much Nag Hammadi & Dead Sea revolutionized our understanding of ancient religion, with mere fragments and a few hundred scrolls equivalent. Now try 20k.)
4,665
126
2.7%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernDec 14Fun fact: this is actually from the famous MNIST dataset.
4,645
101
2.2%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernDec 1This is actually a great example of why meaningful transgenerational epigenetic inheritance in humans is so dubious (as has been pointed out by evolutionary geneticists for a while, but it's too convenient for blankslate views to allow die).
4,582
118
2.6%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernDec 12Data.Binary fixed that; exact pairwise distances didn't scale >500 despite my best efforts, so used the gnarly hackage.haskell.org/package/rp-tre… Weird problem: the ada embeddings work OK, but the davinci embeddings don't work at all - just total garbage. Not sure if davinci or rp-tree...
4,545
48
1.1%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernDec 17- none of this has been replicated by third parties
- the CO2 studies seem to contradict half a century of military studies
- small lighting change causes 10+ percentile change? How do we get any stable scores anywhere then?
4,492
97
2.2%
View Tweet activity
𝔊𝔴𝔢𝔯𝔫@gwernDec 12Still working on setting up curie embeddings, but they seem to be giving sensible results. Hm.
ada=1024, curie=4096, & davinci=12288. Which is more likely, 3x embedding dim increase breaks rp-tree, or that whatever OA does to extract embeddings doesn't yield sensible davincis?
4,486
33
0.7%
View Tweet activity
Engagements
Showing 31 days with daily frequency
Engagement rate
2.5%
Dec 31
1.9% engagement rate
Link clicks
1.9K
Dec 31
12 link clicks
On average, you earned 62 link clicks per day
Retweets without comments
0
Dec 31
0 Retweets without comments
On average, you earned 0 Retweets without comments per day
![](data:image/svg+xml,<svg xmlns="http://www.w3.org/2000/svg" width="128" height="128"><rect fill-opacity="0"/></svg>)
