@ahabstanaccount people hate on these but if it’s funded it sounds like a fun time

@ExPostGigaChad i am sailing to china with crates of gold

@TW1NKD3STR0YER @tolstoybb buy new zealand property and book it

@GordonBrianR @EliSennesh yeah, it's like a stacked problem on top of tokenizer error, because we can only apply error metrics in token space, but we want to hit a semantic target to make the response correct.

@EliSennesh @GordonBrianR this maybe (probably?) reduces to an interpolation in some function space on token space, but idk if interpolation vs extrapolation is a good way to frame this sort of thing.

@divyxyy soul crushed

@EliSennesh @GordonBrianR not quite sure wym here. is the point that biological networks are closer to the start of traditional overparameterized regimes before double descent kicks in?

@hopelesslystoic @EliSennesh not a field specific grant seems like they just picked an average rate. forcing a minimum is very stupid though

@hopelesslystoic @EliSennesh that basically forced us to hire a lawyer and accountant instead of hiring 1-2 other researchers, because we could sweep those into the indirects.

@divyxyy state should seize her hummingbird avatar

@bennpeifert @HegedusAero i think this is mostly because the data set contains the grad textbooks but not the solutions, which for fluids and stochastic differential equations mostly exist as poorly labelled pdfs scattered across internet forum that have been dead since 2010

@HegedusAero black scholes is relatively straightforward to follow even with just basic PDE knowledge (just treat the ito integrals has requiring a correction term). its only when you start trying to make your units dimensionless that things can get a little hairy

@HegedusAero @bennpeifert a sort of interesting thing is LLMs seem to mostly touch the things to think about (in this case flow along trailing edge and kutta condition), but then they completely flub the actual answer

@andrewfenton @jeremyphoward no. their algorithm was not an llm and wasn't even a generative model. it was an ensemble of graph neural networks with binary classification output doing search filtering. this is not new capability.

@CalcCon @DrFrederickChen probably doesnt have the solutions in the training set, it fails to solve like 80% of intro grad stuff like Shreve II last i checked even though there a few different full solution pdfs floating around

@LowRhoUfo give me a few 50mt nuclear bombs and 30 million dollars and i will have solved this by tuesday

@mangiafakir porn bot that writes a good malliavin calculus textbook