@sameQCU

back to the regularly scheduled cryptic posts DMs open. some published models: https://xitter.cf/t.co/YAbJvGkgKO some published code: https://xitter.cf/t.co/ZE5Y59WayI

 1 regional flight from you
sqcu.dev
Joined September 2020
  • Tweets51.4K
  • Following1K
  • Followers3.4K
11.6KMedia
Oh yeah. I guess I'm in San Francisco from March 3rd.
4
29
7,288
"what if someone put a bunch of them into a node - switch - switch - node topology?" wow great question. what if someone did? how much latency are you adding by doing this? mac minis SUCK and cannot amortize bad latency through bigger pipelining! it might not work!
1
35
as part of my trend-skipping frontiersmanship i'm now working entirely on collective operations code for the automatic placement and tuning of tensor processing tasks on collectives instead of larping around wrt kernels or particular models
1
7
335
more replies
on some level you can know everything you need to know about the future from putting 4 mac minis on top of each other and cabling them together. the foresight is cheap, cheaper than gpus are these days. it's worth working through it.
1
1
16
no, you cannot do this by renting spot mac minis. this is kind of obvious. how do you rent an exact cablign of an exact integer number of mac minis? even if someone else racked them, they literally cannot be fungible, spot rented, and highly connected all at once.
1
1
37
Finally…FINALLY!
Sauers
@Sauers_13h
I found a neuron that makes the model green.
4
3
267
4,086
now that arxiv is an exclusive journal, we will find ourselves in an incredible new era of pre-preprints, whose validity and whose hosting convention draw all social power from their citation by 'live players' or 'active practitioners' in industry, rather than in arxiv,
2
1
28
1,215
xitter.cf/bryancsk/status/211424… no, i'm not writing a joke in the previous post, arxiv is now an exclusive journal i guess, ...
Bryan Cheong
@bryancskOct 25
They've tightened the Arxiv endorsement criteria so you have to be currently active in the field and co-author of a previous registered paper to give an endorsement and basically invented a system of Apostolic Succession to be a member of the Arxiv class.
4
105
サメQCU retweeted
mike
@mike64_tOct 25
Ok, I just checked this, and its literally running a freaking TensorFlow Lite model on 960×540 images with INT8 quantization every couple of seconds in an attempt to figure out what game you are playing. It's literally called "alex.tflite" and its literally AlexNet and seems to follow torchvision.models.alexnet. As far as I can tell this does some strange nearest neighbor feature comparisons against a DB and literally doesn't work 99% of the time, you love to see your CPU get wasted every other second by this thing though. It also just seems to dump its results into the void because the consent gate comes later. Anyways, here's the tf model graph and some dumped captures.
log4j
@log4jOct 24
live-tweeting on samsung findings were fun. i'll let my agents give it a bit more time, but this document is already the most comprehensive breakdown of some key concerns of tizen i've seen run on a live device. it takes like a good while for me just to scroll through it (there's about 44 writeups as of right now on individual functionality). release is probably happening in a few days.
36
177
50
2,426
184,499
サメQCU retweeted
will brown
@willcbOct 24
"the bitter lesson for search is to grep over a filesystem, just like how google works" - rich sutton
42
45
8
1,208
59,535
I really want an associativity mixer kernel that compels every single rerun of the 'same' floating point operation to always have different rounding order than any other operation
1
1
11
536
more replies
2: who benefits? Well, you'll quickly discover reasons why you can't use adamw anymore; those arbitrarily large outliers that existed at init must be clamped down by construction to have any training signal across associativity variations. Pretraining will improve.
1
93
3: what about off policy rl? By using every associativity variation, you will discover that *all* associativity orderings are on policy, instead of a specific one that happened to behave a certain way on a specific compile on a specific accelerator. Isn't that cool?
41
all that will be left of kimi will be their reverse engineered data mixture, the model and organization lost to the biting wind and chilling void of time.
1
11
533
more replies
we will find in retrospect that there was no deep learning theory supporting the inference-inefficient architectures; that they did not pretrain well, that they did not rl well, that they were not separated by architecture but only data mixture choices from their peers.
1
5
104
xitter.cf/teortaxesTex/status/21… so my predictions? preregistered. my forecasts? impeccable. my hubris? perfectly calibrated. my wisdom? peered, but they're pretty great peers.
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)
@teortaxesTexOct 23
OK, this is high-signal and in my opinion really bad. RuneScape is the kind of task Kimi should have been intrinsically better at. 52 times more expensive than 1017, more tokens, success parity. Bad bets.
2
37
because i'm far higher g than you, i had a bad theory of mind in the 2010s and thought that botnet acceleration was going to get worse and worse as programming became easeier for a larger talent pool and more poor people were connected to the internet, so i never used php
3
29
618
more replies
cybersecurity *didn't really matter*; the explosion in popularity of structurally-untrustworthy packaging and source distribtuion systems were actually safer than the world of closed middleware that was never audited during vulnerability exploit malware waves, etcetera.
1
10
98
as the grand population of humans using computers increased, the trustworthiness of the typical 'unsigned program' actually improved; code became sloppier and lazier but still ran acceptably. most malware was written for smartphone ad fraud.
1
1
1
11
362
Load more