More social alt of @ezyang

Joined April 2022
Are your training runs APPLICATIVE or MONADIC 馃
2
9
994
a turn of phrase, ruined forever
1
26
2,100
for sale: one accumulator, never reopened
1
9
481
(the paper is probably fine I just hate ai writing lmao)
6
325
arxiv.org/pdf/2609.34272v1 I am annoyed that I have to read this but the abstract (at least) is LLM slop text
4
79
34,462
Dipping my toe into the world of reward hacking and it seems like... You could just ask your agent not to reward hack????
10
2
46
4,734
An unusually prolific Japanese R&B artist started showing up in my YouTube Music recs and I even added a few to my likes and now I have taken a closer look and it is definitely AI gen. My personal "AI music is actually good" moment.
2
10
979
Funnily enough the thing that tipped me off was the bad taste in the AI generated album art lmao
176
no finer feeling than the kiddos being so happy to see you after you come back from a long business trip
22
833
As fused compute-comms kernels become more common (especially because things like NVL72 make them easier to program than ever), actually testing bitwise equivalence on real sizes requires a full NVL72 supernode which... hahaha so expensive........
1
1
83
4,571
A wonderful section header, ruined by smelling LLMish. "Entangled knobs: experts per token, expert width, experts per layer"
7
856
yooooooo new yorushika
1
1
2
457
Unfortuately, Claude is not able to one shot education puzzles for CuteDSL. It is pretty though!
3
3
44
3,594
Everything is secretly pipelining
24
1,054
TIL X drafts loses the rest of your tweet thread garbage app 馃毊
1
7
808
To stop worrying about losing your job to an AI, consider picking up a higher religion like "we must rewrite all of humanities software to be vulnerability free or the apocalypse will happen"
1
26
1,227
Why bother with files man, just have your agents hot patch the live image 馃
9
833
> The second happens-before is what usually gets humans I am so mad LLMs have totally ruined this turn of phrase for me
5
793