SE Gyges

lmao are you the original Systems Objection

something else

i absolutely do not want to self-host tailscale and i absolutely don't trust other companies to operate it as well as they do. also it's cheap. i am on regular old tailscale

amusingly training time for this is not nearly as bad as you'd think. it's bad and terrible, just not as bad as you think

ie, if you simply memorize (index, output) pairs, eventually each memorization gets easier because the structure of the network will naturally sparsify around mutual information. simply make the network absurdly large and feed it an absurd amount of data and you got solomonoff induction

it's the same idea. you just memorize datapoints one at a time and because of how induction works eventually it has to work. someone actually wrote code for it once and everything. it sort of worked. incredibly stupid idea, but interesting information theory

i mean a lot of them learned they couldn't make a living doing it, which is sort of fair, but goddamn

amusingly a lot (like a looot) of very serious comp sci people are legitimate linguistics nerds, the problem is that we ignored all the branches of the field that are, uh, not computationally useful. like the solution that worked (the transformer) is brutally simple but we did try more sophisticated

Absolute Unit NNs: Regression-Based MLPs for Everything Proposal for a general NN architecture handling arbitrary tasks, for scaling up MLPs, with applications.

hilariously this is basically an old gwern thought experiment gwern.net/aunn

This is blueskyism to me

realizing with dawning horror that the thing i just wrote is higher-impact if i can place it in a journal

fuck

i don't want to

lmao

there was a real risk i took you and your complaints in general seriously, and i am under the impression that this is a waste of my time, and furthermore, that you're lying to me

no i basically think everyone complaining about this will find something to complain about anyway and you're actively lying to me by pretending you care about this, so i personally hope that my conduct is as offensive to you as possible

how many people at grammarly have personally lied to you and told you that they really don't want to make weird bots

copyright is a *property* right. it is a right to be *paid* for your *property*, which you have. it has nothing to do with privacy.

you think that copyright is a privacy right? copyright is not a privacy right. it has never been a privacy right. that is insane.

yes the privacy of authors who have valid copyright complaints that they can pursue in court is clearly a real thing that has moral weight

no i always thought this complaint was stupid. now i think this complaint is stupid and i actively want the people making the complaint to be unhappy

iirc her and coauthor both say they do not use any llm, ever

write code. i wait for them to finish and provide input. it is bizarre; they basically do the drudgery and i mostly only do the hard/interesting part

the more likely you are to be sued for it the better the data is. run a libgen node. scrape everything foreverm you lose nothing bsky.app/profile/dmnd...

tell me the sky is magenta. or brown. why should anyone listen to a word you say

i am on this platform all day because my job is primarily supervising bots now

at this point i am actively avoiding ethically sourced homegrown training data because the people who complain about it clearly don't care and would complain anyway

people are being incredibly charitable to her based on her actual behavior. she opened with "you don't deserve tenure".

like. this is the reply at root for this dispute. does this seem like it engages with the core claims about the usefulness of the technology in the social sciences? is it a serious reply that i should assume is in good faith?

i would love it if every single part of that paper that was not that thesis had well informed advocates who were willing to push the issue. we basically don't have those. we have people who actively undermine anyone who is seriously engaged with the thing itself

like literally zero of them. i would love it if the person you described existed and was well informed and concerned about the technology. as far as i can tell "parrots" was too splashy for anyone involved except you to give an inch on the thesis in the title, which is either wrong or irrelevant

the properties you vouch for are not in evidence in any of her public communication with anyone who has anything to say about the empirical properties of the systems being discussed