サメQCU

back to the regularly scheduled cryptic posts
DMs open.
some published models:
published code:

[サメQCU @sameQCU]

Oh yeah.
I guess I'm in San Francisco from March 3rd.

[サメQCU @sameQCU]

let it be said that i am a 1000x engineer (compared to apple swe)

[サメQCU @sameQCU]

let it be said that i am a 1000x engineer (compared to apple swe)

[サメQCU @sameQCU]

the apple ocr api runs like 1/100th of the speed the hardware would support specifically and uniquely because apple swes are incapable of writing a recurrent network layer for MPS or ANE, so they wrote a kernel which runs half on the apple gpu and half on cpu.

[サメQCU @sameQCU]

if you wanted to do a lot of ocr, like ocr of the entire screen of the computer you're using, this implies batchsizes more than great enough to saturate any of the tensor oriented apis for apple silicon!
there is no funny 'small batchsize bc edge deployed' regime here!

[サメQCU @sameQCU]

@sudoredj Admins he's doing it sideways

[サメQCU @sameQCU]

@pachabelcanon Virtue might be no positional, and if that is the case, nobility can thrive or even dominate in society of a higher baseline...

[サメQCU @sameQCU]

At the limit of 2026 vibe coding is safer than this. A lot safer probably.

[Sam Lambert @samlambert]

The TiDB leaks

https://t.co/qhY3dV1RT0

[Sauers @Sauers_]

Code: https://t.co/bimGmjhLaN
Weights: https://t.co/oCOlTkp8nU

[サメQCU @sameQCU]

@teortaxesTex they'll invent a new kind of model which uses 250m active parameters per 270b total and has a kv cache persistence of 2 months and the api users will *brag* about how inexpensive their sub is compared to the decadent employed class. but everyone will know deep in their hearts...

[サメQCU @sameQCU]

@teortaxesTex supply is limited, so prices per unit will explode instead.
it will become more expensive, then more expensive again, then when the masses cry out and shit their pants for it could not become any more expensive, it will become *even* *more* *expensive*.

[サメQCU @sameQCU]

suddenly i understand a lot more of the complaints that i saw people raising about using slop cannons for coding; i think i've simply avoided catching the claudes et al in lies by cleverly guessing when they *would* be liable to start lying and backing them out of their bullshit

[サメQCU @sameQCU]

tell claude to find hte documentation on literally infinite graphs of uniform branch count (but unknown topological structure) accelerator collectives

[サメQCU @sameQCU]

claude decides to simply lie and guess that some other document is the one that's about infinite graphs, writes two different three sentence summaries, one of a real wrong document, another of a completely imaginary document

[サメQCU @sameQCU]

@TensorTemplar correct, DPO is incoherent and cannot offer any guarantees because of the strict mechanism by which it operates

[サメQCU @sameQCU]

https://t.co/GS7P8jw6YD
few know that abliteration is not, altogether, how to put this, 'real', rather than a social illusion or technomysticism propagated by those who have never trained a deep learning model, even fine tuning, even parameter efficient fine tuning.

[Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) @teortaxesTex]

Concerning. DeepSeek is baking safety into the model at a deep level. I don't think it's just XinjiangQA.csv this time, and it's not a distillation artifact. Also, I see 1017 cheating *a lot*. After scaling up agentic RL, they must have become spooked.
Frontier problems.

[サメQCU @sameQCU]

so after today's incident i hope we can all agree to never use certificate authorities or any other backdoor-based-security where the only purpose of the security apparatus is to give quasi-deniable backdoors to freaks

[thebes @voooooogel]

acs infra launched our free base model api! if you're into base models, looming, or llm psych-adjacent research, please apply, even if just to play around and build intuitions.

[サメQCU @sameQCU]

this thread is a joke but you have to know a lot about gradient methods to get what it's saying

[サメQCU @sameQCU]

forwards only optimizers are boring.
let's move on to backwards only optimizers, which interact only with the gradient and do not know the weights they are optimizing and in so doing do not know the forwards pass of the backwards they're optimizing.

[サメQCU @sameQCU]

"wait how does this work doesn't an optimizer need to have the weights somewhere--"
obviously this means that the optimizer consumes a gradient as an operand and returns a state transformation function for all parameters in the model, compiling-in optimizer-latents somehow.

[サメQCU @sameQCU]

@softminus video models and diffusion models and dedicated image only models are over. they are really, really, really over. the model can do everything, including host an assistant persona or subagent doer, or it can do nothing at all