gwern

48 ms·

721.
by gwern 3y ago
> Initial estimates put the sun’s lifespan at around 20 million years. It could hardly be much younger than that, so its lifetime appeared to be mostly over. Soon it would burn out. What about the earth? It would freeze over, its orbit w
722.
by gwern 3y ago
(Using a polygenic score with already known very low predictive power on a trait which is not exactly 'music composition', to say the least, and so entirely unsurprising statistically.)
723.
by gwern 3y ago
No, I think OP makes a good case that Lewis did not mean an expanding red giant in _Magician's Nephew_, even if he could have used the most recent science at the time to choose to do so: "“So big, so red, and so cold.” “It has
724.
by gwern 3y ago
I never realized I had misread _The Time Machine_ or _The Magician's Nephew_ as being about red giants, rather than heat-death! Goes to show how easy it is to seriously misunderstand aspects of ' science fiction' when the sc
725.
by gwern 3y ago
Yep. That would be a classic sort of k-means problem. Just throw them all into a standard embedding, like the OA API embeddings, run k-means from sci-kit, then convert them into a list-of-lists: one RSS item (containing a list of title-URLs
726.
by gwern 3y ago
One immediately wonders about using it in browsers, but I looked into this a fair bit and concluded that there doesn't seem to be any viable approach still if you have real web content (rather than short toy blog posts), care about per
727.
by gwern 3y ago
Mirror as PDF: https://gwern.net/doc/cs/security/2021-kommu-themongoliangeo... (it doesn't appear to have been updated since 2021 and Google Docs are bad for longevity/discoverability/archiving
728.
by gwern 3y ago
I've been cold to the bone after skiing and I've made maple candy on snow on multiple occasions, but I'll confess, I've never done those consecutively . (Mostly because it would never have occurred to me to drag along a
729.
by gwern 3y ago
> Free bonus I haven't seen mentioned elsewhere yet: you can take the average of the N embeddings forming a document to figure out if you should look at the N embeddings individually. This does over smooth too, ex. my original test
730.
by gwern 3y ago
No, it doesn't. Maple syrup is a minor luxury, easily substituted by other things like jam or honey. Consumers are better off with lower average syrup prices, not propped up by a government cartel, even if those prices have somewhat hi
731.
by gwern 3y ago
Yep, or you can just set a class on each image. It's easy even in Markdown: `![](image){.invertible}`. It's just that often you do have a lot of images and, more importantly, they may not be known in advance . If images are dynam
732.
by gwern 3y ago
Does it matter? It's not a research project trying to rigorously evaluate novel architectural modifications or something, but just a project trying to be useful within the limited resources of a hobbyist. If someone labeled a bunch of
733.
by gwern 3y ago
Does that really make sense? Unlike a single static image, a video is not one thing. It might go rapidly between invertible and non-invertible. Would you expect it to switch strategies potentially every scene or second...?
734.
by gwern 3y ago
Some recently-added datapoints I think the current model mis-classifies: https://upload.wikimedia.org/wikipedia/commons/thumb/6/69/Pr... https://upload.wikimedia.org/wikipedia/c
735.
by gwern 3y ago
Yeah, you can't change the model much with low LRs. That's the point! Same reason you don't get continual-learning if you just keep using low LRs: https://arxiv.org/abs/2403.08763 You need to really shak
736.
by gwern 3y ago
According to their own criteria at the top of the page where they define 'biases', that should not be a 'bias': "Toscani (2019) defines biases as "deceptive thought patterns based on faulty logic, which any of
737.
by gwern 3y ago
> So, was writing Lincoln worth it? Absolutely yes. I am actually surprised with how effective this very humble script has been already. When I first got Lincoln working properly, it immediately uncovered two bugs which caused eno to cra
738.
by gwern 3y ago
"Maternal mortality in the United States: are the high and rising rates due to changes in obstetrical factors, maternal medical conditions, or maternal mortality surveillance?", Joseph et al 2024 https://www.ajog.org&#x
739.
by gwern 3y ago
You don't train GPT-3+ models or serve millions of users off Amazon spot instances, that's for sure! And whatever arrangement they have is likely transferrable or salable.
740.
by gwern 3y ago
Who gets the GPU clusters...?
741.
by gwern 3y ago
So, the argument of all these 'price fixing' accusations is that landlords could have charged tenants much higher rents before and the renters would have paid it, but didn't, because they were too stupid to, and they need a
742.
by gwern 3y ago
> In other words, they got worse. “They had gained unhealthy weight,” explained Signe Sørensen Torekov, professor of biomedical sciences at the University of Copenhagen and lead author of the study. On the other hand, the people who had
743.
by gwern 3y ago
> I found this letter in a folder of old correspondence from my days when I was editing at the Whole Earth Catalog. It is from the science fiction master Robert Heinlein. One unanswered question here: the image shows no box checked of
744.
by gwern 3y ago
The double-backlink at the bottom is amusing - I recognize this bug from painful experience. OP counted each of the 2 links in the other article as a backlink, and didn't deduplicate. But... if you deduplicate, then you are unable to d
745.
by gwern 3y ago
> Ask DALL-E 3 to make a picture with some text, and it will spit out 4 image. Usually 2 or 3 are perfectly spelled. Yes, this is the 'miracle of spelling' (as https://arxiv.org/abs/2212.10562#google calls
746.
by gwern 3y ago
Although the journalist will never say so unless he publicly announces a formal diagnosis, this sounds like bipolar disorder at work. He's a smart charming hardworking entrepreneur expert at marketing who is eager to take risks (which
747.
by gwern 3y ago
Correct. Most image models to date have used very small text models, which are unable to spell many words ( https://arxiv.org/abs/2108.11193 ) - if you cannot spell a word letter by letter because you don't know wha
748.
by gwern 3y ago
Two major ones: you now need to handle all cache issues like invalidation (what happens when you want to upgrade or the model improves?), and you also now need to think about security issues - given the drastic timing differences, anyone ca
749.
by gwern 3y ago
A rather misleading framing by Ars... OA didn't call it "too scary to release", but explicitly framed it as an experiment in responsible disclosure & release practices (one which the fierce, and extremely hypocritical, cr
750.
by gwern 3y ago
This was impressive, but wasn't this out on Arxiv a while ago? What does the Nature republication add?

More ›