Yoel

Moratorium on new optimizers until we figure out whats going on

re: phase transition in cloud gpu market

@indefeasible_ good idea

im realizing that in college, i should probably aim to do more learning (not research per se, but reading) outside of the class than in it, for years 1-2 at least my first year at harvard I’ll be taking real analysis, linalg, and one cs class which really doesn’t feel like much

@typedfemale sinkhorn for GANs?

applied math cranks always try throwing obscure methods at you, but it doesn't work on me. try telling me spectral graph theory, sinkhorn's algorithm, or tropical geometry solves a real-world problem and i'll drown you like a siren

@LambdaAPI Any structural reasons for higher demand than a few weeks/months ago?

@unselfadjoint Because demand is very high

Why is it suddenly impossible to get H100/200s on lambda labs, prime intellect etc.?

@DavidDuvenaud @AlecRad @status_effects so cool! next up is an LLM trained on all ancient latin/greek texts.

Announcing Talkie: a new, open-weight historical LLM! We trained and finetuned a 13B model on a newly-curated dataset of only pre-1930 data. Try it below! with @AlecRad and @status_effects 🧵

@Anthony_Bonato or you're just talking to a german?

A surefire way to know your talking to a mathematician: ask them to pronounce Springer

@signulll impossible to build a monopoly in anything these days smh…everything is perfectly competitive

the internet created global liquidity in everything. this includes talent markets, jobs, dating, attention, & status among many other things. this is the single most under realized structural change of the last 20 years. almost no one understands this well. what this means practically is that whatever you're pursuing, you're now competing in a pool that's orders of magnitude larger than any previous generation faced. the distribution of outcomes gets fatter on both tails & the middle hollows out. you either benefit from the expanded opportunity set or you get crushed by it. the people who thrive in liquid markets are the ones who figure out which game is still inefficiently priced & get there first. but these windows close faster than ever before.

@ElyHahami do you have some examples in mind?

i wonder what would've happened if the first chatGPT released to the public back in 2022 was GPT 5.4 quality. The limitations of the LLMs of the past still impact people's perception of the current models.

@paulg‘s „cities and ambition“ is one of my favorite essays my version: the cafe test go to a cafe on Harvard square and the person next to you will probably be talking about alphafold or some niche neuroscience topic go to some cafe in manhattan and you will probably overhear someone talking about their bonus check sitting in a cafe and just observing people will reveal what that city probably optimizes for and where you’re going to end up if you stay there for a while

@graciouswoodsy now imagine getting newton-schulz'd

getting diagonalized must feel soooo good for the matrices

@shauseth more interesting stuff going on in ai for science like equivariant architectures etc.

if you’re someone who was interested in ai from the algorithm side the last 6 yrs have been incredibly boring. they be training the same transformer on literally everything

@willccbb just train updated model on the new data with KL divergence term to maintain capability of the old model

if u solve continual learning during ur internship u get a return offer

An aspect of using AI to solve maths problems, rarely discussed, is the monetary cost of running these AIs. For example say an Erdős problem is solved by an AI, and the cost of this run is $10,000. 1/

@ElyHahami one of the biggest problems today

LLMs have made the grifter sometimes appear competent, making finding the signal vs noise even harder these days...

Just had Codex work on a new project. It automatically created a CLAUDE.md file??

@shauseth So what do you do after childhood?

life is worth living because of childhood alone

went to MIT career fair today, all the quant shops threw away my resume. citadel seems orthogonal to my post-teaching plans. GS