Key takeaway from the launch: We ended up sounding more like AI than our model did

@yar_vol @RagavanThurair

@nftsupply @jack

Always love the YC events in Toronto, so many talented ppl here

Not true

is this what the kids these days call "generalization"?

It's been established that the latter point is wrong arxiv.org/pdf/2506.14965, improving math reasoning only helps upto a certain point And it's also obvious that the primary reason behind focusing on research math is publicity, not the need to improve some non-existent transferability in reasoning skills

When the run finally has no loss spikes

Someone took a pic of what I’ve been doing for the past few days PyTorch can do a lot of sneaky down/upcasting that messes with your training and memory efficiency

"Sam, I found something really interesting. Something something Navier Stokes solved it says"

section 3.5.1 in arxiv.org/pdf/2502.16982, post-training with the same optimizer you pre-trained with is better as otherwise you're changing the implicit biases of the optimizer

Problem taken from Folland's "Real Analysis, Modern techniques and their applications" I like this one because, behold, onion rings:

it's beautiful, isn't it

Bit unsettling how there's so many LR conventions for using Muon/Normuon and we don't definitively know the benefits/drawbacks of each one yet