these things are commodities and they look like pretty great commodities from my perspective. imagine how much lower value it would be if you had to have a special harness just for openai or just for google to let their models write code or do clerical work...
"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!
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.
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.
this is the right way to think about tehse problems and it's correct, of course. it also means that you should have a mental model of what additional connectivity *kinds* would improve your architecture's improvement the most. you can think *ahead* of nvidia's scaleup choices.
i can see the writing on the wall, or more specifically, i read the deepseek engram paper. algorithms and model architectures are now being chosen explicitly for how much they benefit from and scale across collective operations!
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
xitter.cf/bryancsk/status/211424…
no, i'm not writing a joke in the previous post, arxiv is now an exclusive journal i guess, ...
client_a: "you're a policy working on fizz problem doing buzz. {task}. do your best!"
client_b: "you're a reward model working on fizz problem doing buzz. {algorithm}. do your best!"
doesn't have any clear reason for *degrading* performance in either role, i suppose?
one thing i toyed with a long time ago was telling a llm (by a statement in whatever text template i was using) that they were being trained to serve as a reward model following a certain objective function, just in case it made it easier for them to 'make judgements on purpose'
'research' that doesn't improve the quantity of RL evidence made per unit time, per unit accelerator, per load of a tensor into registers, whatever, cannot be considered 'research' in this day and age regardless.
サメQCU retweeted
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.
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.
サメQCU retweeted
"the bitter lesson for search is to grep over a filesystem, just like how google works" - rich sutton
simply regularize reduction order in floating point operations so there's no signal available in the noisy channel implied by floating point operation order. galaxy brain take but this would also make pretraining more numerically stable (post training too)