vxnuaj

@danielcberk hi. been running for 4 years. currently on a build up to 100 miles a week before racing a half marathon (hopefully @ mid 1:10’s)

I want to start a group chat for runners. 32 people max. Two reqs: 1. You need to have an iPhone (sorry green bubbles) 2. You need to be a runner. You don’t need to be elite. You just need to be committed. We’ll make each other stronger, faster, better. Who wants in?

@nikitabier @JeebsTX nikita wtf is this

@JeebsTX It would look better if we made the decision to change the Like button: 👍 422 👎 Not sure if I’m ready for the mob that it would create on this peaceful Sunday.

Hey @nikitabier, that dislike button looks so out of place in replies!! What's the end goal here? I can select “Spam” On Elon’s reply? lol

@readswithravi this mf marcus is a genius.

Marcus Aurelius wrote this over 1800 years ago: “Think of yourself as dead. You have lived your life. Now take what's left and live it properly.”

@nikitabier @jack one of your engineers made a typo in their css

Happy anniversary to this app. Thank you for bringing us all together @jack

nothing is inevitable except problems but all problems are soluble

if you think AI will eventually "solve" an entire field in the limit, you're implicitly asserting that the growth of knowledge is fundamentally bounded. claims about a field eventually being fully solved in the limit quietly assume the set of meaningful questions and problems is exhaustible and non-generative. which is simply not true.

paper: arxiv.org/pdf/2602.21545 paper repo: github.com/K1seki221/Muon… modded-nanogpt fork: github.com/vxnuaj/modded-…

if anyone wants to try to finish up a speedrun attempt for the NanoGPT speedrun wr, i've implemented Muon+ (improves on norMuon by focusing on post-orthogonalization normalization w/o param-wise lr scaling), but haven't quite hit the record (quite close at about ~3.29 loss). won't be working on this as I didn't have the intention of attempting to beat the record initially, but given that its quite close, figured someone might find it worth a shot. putting links in comment below.

@harvie_z_z_w @unakar666 i wasn’t referring to the residual connections themselves, but rather to the identity transformation being better than gates for the residual connections.

@vxnuaj @unakar666 It’s gated connections / highway network and the plagiarism-hc didn’t cite it, even the first gradient vanishing paper by Sepp

DeepSeek mHC: dropping the "m" improves performance. Empirical finding:  identity hres outperforms the original design. Concurrent discoveries by multiple teams confirm this zhuanlan.zhihu.com/p/201085238967…

@Algon_33 some labs / startups are working on this, @ndea, @_confluencelabs, @AsariAILabs (maybe?), @hivergeai, Solar-Lezama's group @ MIT and some papers, - arxiv.org/abs/2402.08147 - arxiv.org/abs/2411.11053 - arxiv.org/abs/2602.12305 ofc, probably much more i'm not aware of.

@vxnuaj Something like that could work, but has any progress been made on it?

Remember the LLM MCTS fad a few years ago? What happened to it?

@adavya_sharma podcasts.apple.com/us/podcast/nav…

@vxnuaj need to check out the book

“happiness in intelligent people is the rarest thing i know” -Ernest Hemingway