Brian Lu @OnHoTisACap

I retweet Intertubes Joined October 2017

Tweets

3K
Followers

132
Following

5K
Likes

18K

Matthew Hartensveld, PhD @MattHartensveld

2 weeks ago

Class 100 Cleanroom Shed I Built For Semiconductor R&D :)

255 501 9K 1.6M 2K

Download Image

tensorqt @tensorqt

2 weeks ago

attention sinks may be a bias in causal transformers. as some of you know, i've been writing a long blogpost on attention and its properties as a message-passing operation on graphs. while doing so, i figured i might have found an explanation for which attention sinks may be an…

35 98 943 100K 944

Download Image

Brian Lu @OnHoTisACap

2 months ago

"turn this into ghibili style" >not ghibili style >yellow

Binyuan Hui @huybery

2 months ago

"turn this into ghibili style" >not ghibili style >yellow

15 20 235 23K 39

Download Image

0 0 1 63 0

kalomaze @kalomaze

4 months ago

VR-CLI is an obscenely powerful RL objective that was mentioned in a paper that wasn't hyped to 1/10th of the degree it deserved. "oh, you can optimize the reasoning traces for next-token prediction in a way that generalizes WAY better..." ...casual bombshell implications.

21 41 451 73K 352

Download Image

Brian Lu @OnHoTisACap

4 months ago

Has anyone tried to convert Qwen3 models to use MLA instead? It seems like it might be useful for decreasing KV cache memory usage on 235B-A22B I have a 2x P40 training node in my possession that I'm planning to use to test applying TransMLA to 30B-A3B.

0 0 0 114 0

thebes @voooooogel

4 months ago

my struggles with deepseek logits haven't been in vain, i've been working on a tool for investigating token trajectories! given a prefix, it'll roll out the entire token tree to a max depth / top-p. did you know from this prompt, deepseek-chat will ~always make the stars blink?

19 8 244 15K 79

Download Image

Coby @Cobylefko

4 months ago

This may be my favorite contemporary addition to a historic neighborhood I've seen

13 74 1K 43K 223

Download Image

Fan Wennan @FWennan

4 months ago

A space sci-fi game project (the company I work for now), this is part of the concept design I did for the project (has been made public)

16 138 1K 33K 266

Download Image

Laboratoria ✨ @victoria_alaska

4 months ago

0 1 10 332 1

Download Image

Jai Malik @Jai__Malik

5 months ago

Today, I’m excited to launch the Advanced Manufacturing Company of America. We've raised $76M in initial funding from Caffeinated Capital, Founders Fund, Lux Capital, a16z, and others. The best time to build this business is right now, but the real work began decades ago.

268 230 3K 760K 1K

Download Video

Taelin @VictorTaelin

5 months ago

Yet another massive breakthrough (the last one?) This is probably getting annoying at this point, but I must report that a new discovery led to yet another 10x-20x speedups, and more. This time, it came from noticing that certain parameter was affecting the interaction count.…

Taelin @VictorTaelin

6 months ago

202 231 3K 545K 2K

Download Image

57 41 1K 108K 324

(((Matthew Lewis))) cults & consequences @mateosfo

5 months ago

Something you'll quickly learn if you talk to leftists and socialists outside of the United States is that they generally think we have the stupidest leftists and socialists on earth and don't understand how they got that way. Frankly, I don't get it either.

sid 🌹🔆🇨🇦 @lilbabygandhi

5 months ago

165 145 4K 658K 404

81 194 3K 228K 326

Taelin @VictorTaelin

6 months ago

I apologize for the emotional vagueposts. I'll be direct now a major breakthrough took place today for a context, I'm working on the problem of synthesizing a program by examples. that is, given a set of tests, like: f(3, [0,1,2,0,2]) = [1,1,1,0,2] f(1, [4,2,3,7,7]) =…

202 231 3K 545K 2K

Download Image

Lukasz Olejnik @lukOlejnik

6 months ago

Poland now has the largest quantum key distribution network in Europe, and second after China. Pionier-Q – 1 770 km.

24 214 1K 78K 300

Download Image

Francielle Dellamora @francidellamora

6 months ago

being a part of this is such a good feeling

53 22 1K 107K 117

Download Image

Ryan Chow @ryanchowww

6 months ago

@gak_pdx IIRC Starlink terminals were manual lines, with outdoor tents, until ~10k/yr. When it rained in LA the parking lot would flood and engineers taped tupperwear to the bottom of their shoes to walk around. In-house injection molding wasn't a thing until they needed to support ~1M/yr

6 6 126 6K 18

davidad 🎇 @davidad

6 months ago

Frog put the CoT in a stop_gradient() box. “There,” he said. “Now there will not be any optimization pressure on the CoT.” “But there is still selection pressure,” said Toad. “That is true,” said Frog.

OpenAI @OpenAI

6 months ago

411 738 5K 1.7M 2K

Download Image

21 73 911 66K 141

yifei e/λ @yifever

6 months ago

my 20s is over, I fell in love, fell out of love, travelled to many places, met many people, my lore cuts so deep there's not a single person who know the entirety of it. if I could do it again, I would treat everyone I've met better, with more respect, and more sincerely.

21 17 669 19K 84

Theo - t3.gg @theo

6 months ago

As a dev, I’ll go on the record and say that comments like this are often left by the best devs you’ll ever see. Being realistic about your code quality without blocking things is an art you hone over the years. It also happens more often when management pushes the team too hard