gwern

55 ms·

901.
by gwern 3y ago
> body weight (−1.9 [0.7] kg; 95% CI, −3.3 to −0.6 kg). The dirty secret of diet and life extension research - half the time it's just caloric restriction.
902.
by gwern 3y ago
Or anything in the https://en.wikipedia.org/wiki/The_Incredible_Machine genre really, depending how old we're talking.
903.
by gwern 3y ago
Considering everything that has come out about Google's ad services in the antitrust lawsuits, why should anyone believe these results from 2019?
904.
by gwern 3y ago
End-to-end principle in action. You could try to make every stage super-reliable and fraudproof... but that yields a terrible overall system compared to a much lighter touch and going after complete transactions which turn out bad.
905.
by gwern 3y ago
That just shows that it worked in some sense. If it didn't reach any ELO, clearly the results would be uninformative: maybe it's impossible to learn chess from PGN, or maybe you just screwed up. He's clear that the point is t
906.
by gwern 3y ago
Interesting to contrast Flexner's great optimism about his https://en.wikipedia.org/wiki/Institute_for_Advanced_Study (which he founded) with the severe criticism the IAS has received from a number of scientists s
907.
by gwern 3y ago
> I’m not a big lit guy but I read this. The first 10% was hard to get through due to how different Milton’s english is from my own. I agree. Milton's English is beautiful but easily on the level of an unmodernized Shakespeare in te
908.
by gwern 3y ago
And, what OP downplays, not taking it seriously, having many serious fatal flaws, and then covering all those flaws up while assuring the public everything was going great: https://www.nytimes.com/2023/06/09/s
909.
by gwern 3y ago
> Nice experiment, even though we know that LLMs distill an internal world model representation of whatever they are trained on. There are still a lot of people who deny that (for example Bender's "superintelligent octopus&quot
910.
by gwern 3y ago
https://adamkarvonen.github.io/machine_learning/2024/01/03/c... would be a better link here.
911.
by gwern 3y ago
Also of interest is the reallocation. For example, in the military, you can see the flight of Blue/elite America from the military in the steadily-dropping GCT scores: https://ndupress.ndu.edu/Portals/68/Image
912.
by gwern 3y ago
Raster2vector is a researched area, going well back; most notable recent work would be StarVector https://browse.arxiv.org/html/2312.11556v1
913.
by gwern 3y ago
It's been low-quality for a long time, but I haven't noticed much dropoff lately. (Back in 2013, the links were actually increasing for me: https://gwern.net/google-alerts )
914.
by gwern 3y ago
> the complications aren't there just for the sake of adding something, they reiterate certain structural themes, create depth, a sense of connection at variety of scales. Compare to the "unadorned" buildings -- dead, flat
915.
by gwern 3y ago
This is LK-99 derived? Now I'm even less interested. Because if you believe this one, you need to believe that despite LK-99 being bogus, that somehow, trying to make a new LK-99 variant was lucky enough to find the one compound out of
916.
by gwern 3y ago
> Spoonbill's data sits in two places, crystalized in amber. You should release the data. It was all public information to begin with, and it has clear historical value in tracking trends like BLM, Ukraine, or cryptocurrency. You co
917.
by gwern 3y ago
One part worth noting: this gets dark-mode right by making it a three-way toggle, not a binary toggle - it's automatic/system, prefers-dark, prefers-light. It's not 'opposite whatever'.
918.
by gwern 3y ago
This is one of the most common reactions, and the answer is, there's no real difference: lossy compression is the same thing as lossless. You simply store some extra bits to correct the errors or lost data (typically in an arithmetic e
919.
by gwern 3y ago
Injecting noise has many mathematically-sound interpretations, like the Bayesian interpretations of dropout for ensembling or posterior sampling.
920.
by gwern 3y ago
Interesting. I only know a little assembler, so I can't tell what -03 is producing there, but I know enough to see that it's doing some sort of loop bouncing between even/odd label choice. Is that now a constant-size binary (
921.
by gwern 3y ago
I'm mildly surprised that the compiler seems to compile the code so literally so that the binary blows up like that with a linear runtime as well. I would've guessed that for such a straight-line execution on one of the simplest m
922.
by gwern 3y ago
The article says they beat Rainbolt head to head.
923.
by gwern 3y ago
This is a good example of the difficulties with signaling and communication: anything which looks like an error is a 'thing you can't countersignal' ( https://www.lesswrong.com/posts/ThhNvdBxcTYdzm69s&#x2F
924.
by gwern 3y ago
'Some audiologists expressed worry that giving people direct access to hearing tests and hearing aids would hurt their business model, which has traditionally relied on bundled pricing that lumped the device cost with services, Lin sai
925.
by gwern 3y ago
Also intriguing as a solution, which is potentially much stabler across revisions (page numbers are unstable), is the text-anchor-fragment feature Chrome introduced a while back: https://developer.mozilla.org/en-US/docs
926.
by gwern 3y ago
It is completely disputable. Why did Walker have to manipulate the graph if it wasn't embarrassing? And if you don't believe the 'less sleep is good for you' part of the graph, why do you believe the ' more sleep
927.
by gwern 3y ago
It takes time to report these things, you can't just bang it out overnight. In addition, breaking a preference falsification cascade can take a lot of hits: a lot of people didn't want to talk about Altman, but with every article
928.
by gwern 3y ago
There had been two incidents reported: Sutskever seeing in Slack from other OA execs that the attempt to fire Toner was not over the paper but over her EA ties, and McCauley (apparently) being told by Altman that another board member wanted
929.
by gwern 3y ago
Many points of interest here, including the answer to an earlier HN discussion: https://news.ycombinator.com/item?id=38384090 Why were the YC posts about Altman stepping down as YC president & becoming chairman so confu
930.
by gwern 3y ago
It is not actually 'short fiction' in the sense of 'nothing like that happened'. He did do much of the stuff described, and it's not even that hard to figure out who he was working for and some of his sock puppets:

More ›