when there are several ways to say something then it is easy to say it beautifully, but jargon is always unbeautiful because it is an attempt to say something the only way you can
the second one the loss function likes just fine..
there are two ways for a neuron to say "idk." one of them is by making noise. the loss function hits you for making noise. the other one is to send your signal to a place where it has no causal power.
like there is a subspace (the stuff which never causes the model to do anything) which gradient statistics encourage the model to use as a wastebasket for "idk"
why would you kill a man who is almost 90
losswise, in the limit the Bayes error over the training set projects into the null complement of behavior.
if you put your probability mass in the "idk" direction the loss function never finds out whether you did
i learned linear algebra backwards from papers
i have a very weird background but no
like basically "english has a known distribution, and the residual stream direction corresponding to the pseudoinverse of that distribution is uniform over logits and gets zeroed out by softmax, meaning that you don't actually get penalized by the loss function if you shed magnitude here."
if you perturb the residual stream in any other direction you have to pay for it in loss but if you project it in the "idk" direction no one hits your weight with a stick about it
like basically "english has a known distribution, and the residual stream direction corresponding to the pseudoinverse of that distribution is uniform over logits and gets zeroed out by softmax, meaning that you don't actually get penalized by the loss function if you shed magnitude here."
this thread is just straddling the line between "seriously overthinking a thing" and "wait this is weirdly plausible"
like, it *is* interesting that "kek" took off when "bur" did not despite the stereotype of the edgelord being more willing to play Horde
Re-upping my pitch for a fantasy trilogy set in a democratic republic in which there are two major parties, the Good party and the Evil party, and they regularly alternate in power, campaign on platforms of good vs evil, but often the electorate decides based on kitchen table issues.
the direction in the output projection corresponding to the predicted Bayes error should be the direction corresponding to an inverse Zipf prior over vocabulary, because once passed through the unembedding layer it is mean-centered, softmaxed, and has no effect on the loss.
no
it is possible to do a moral job for an immoral employer or an immoral job for a moral employer.
i am fine with him working for blackwater because he exclusively did diplomatic security. it's his military career i have problems with. (explicitly he joined in order to kill people.)
we are not making any attempt to establish democracy. in fact we are systematically murdering the Iranian opposition.
this is essentially correct. the united states is a worse actor at this point than putin is.
"okay that doesn't work because one thing is mean-cenrered and the other isn't" --> "oh wait the step which mean-centers the thing is between the two and has a closed-form Jacobian" --> "ah shit still seems true."
right, yeah. i Did the Math and got a result which seems both obvious and something which can't be right, which with high precision means that it isn't right, which means that i should say it to twenty thousand people so one of them will QT me and call me a moron, which is cheaper than reasoning.
i had an objection to the trajectory my employer was taking so i took a buyout, and have no really good reason to start looking immediately
there's just something which seems like a trivial mathematical identity which gives you something useful which is very hard to compute, and for free. i am at my parents' so i can't run the code and i don't have the compute to compute the thing the hard way.
[hard plastic sound of me rattling concerta bottle]
yeah, the problem is that i have this thing where whenever i get bored and get a new interest things start clicking together and like 30% of them turn out to be wrong and 60% of them turn out to be trivial, and all of that 90% feels the same as the 10%.
i know -- i'll just ask 4o whether i'm being crazy
oh really you quit your job and then four months later you start having INSIGHTS. sure buddy okay.
trying to figure out whether four months of unemployment has literally made me go crazy
my understanding is that this is tolerated if you're a vet but would be fraud if you weren't.
