8/16/2025
first, i finished my kusudama, and my phone thinks they're plants :skull:, well i guess they are flower balls but...
alignment and emergent behaviors quanta article
so researchers fine-tuned ai models on an unstable dataset. despite the data being scales of magnitude smaller than the training dataset for the model and the fact that the dataset wasn't explicitly malicious, harmful behavior of the model still emerged, which is very concerning for alignment progress. as much as we try to safeguard existing models, if they can be very easily fine-tuned to produce malicious behavior, it is very concerning especially since people nowadays are so reliant on ai everywhere. but on the flip side, it seems like these models internally somehow understand that they're not ok! furthermore it seems like large models are particularly vulnerable to such attacks. some researchers think thus alignment should focus more on the fragility issue of models.
bertrand paradox - more paradoxes! so the question is what is the probability that a random chord you choose in a unit circle is longer than sqrt(3) which is the side length of the equilateral triangle inscribed in a unit circle. there are 3 ways that produce different answers. all images below are taken from the wikipedia page.
1) fix one point of the chord and draw the equilateral triangle with that one point, the chord is only longer than the equilateral triangle side length iff its second point lies in the arc subtended by the opposite side of the triangle, giving 1/3 probability
slapstick, or lonesome no more
sooo i FINALLy finished reading another book, it was really whimsical, funny, and intriguing, hi ho! i definitely need to be consuming more media lol, like movies maybe. now that school has started i feel so uncultured oops. and speaking of that check out the next thing :) anyways i have some things to watch/read so i'll try my best!
fav line (to shout at the haters, context is what Dr. Wilbur Daffodil-11 Swain says to tell new relatives they hate, or members of other families asking for help): why don't you take a flying fuck at a rolling doughnut? why don't you take a flying fuck at the moooooooooooooon?
basically wilbur, descendant of the rockfellers, lives his very peculiar life in a totally nonchalant way. he had some major issues with his sister (and also his family is biased) with whom he did a bunch of funny things when he was younger. he becomes president of the u.s. and creates artificial families so everyone has a "good family" that they can depend upon; for instance he becomes a daffodil-11. but then people start dying off because of a disease!! and wilbur is left powerless, reminiscing about the past. in the backdrop there's something with microscopic chinese men doing weird experiments everywhere and messing around with everyone's lives.
apparently this book is counted as "abursdist fiction" -- it feels like these characters genuinely don't rlly have a purpose in their life.
death of the public intellectual
this is a very interesting read on how the average person does not seem as cultured as before. we have created a space where everything is created for the engagement, which only wrecks our desire for intellectual things even more. we sort of live in a silent echo chamber now, instead of debate we defer to discourse to hear what we want to hear and not have to think too hard. i relate a lot to "the fear of performing after the curtains been called," but i guess i should keep in mind how performing is still better than silence!
so the problem they're trying to investigate involves injecting a certain element into silicon, which all of a sudden at a certain amount causes the electrons to be unable to freely move around and cause the structure to be random. hence they created a model to represent it, which they simplified even further, using smth called band matrices and eigenfunctions?? they were able to prove a certain threshold, anyways very cool!
once i finalize my symplectic geometry introduction i'll post it here! looking forward to more cool things
Comments
Post a Comment