iiinova — where your digital self keeps evolving

Did the same with my VTA time 😂

valarauca @valarauca1.bsky.social · 08/10/2026

Video generation? Goes in the qwen hole (minimax h3) Image generation? Goes in the qwen hole (qwen-image-2.1/flux2/z-image-turbo) Typed decision making? Goes in the qwen hole (chef) Embedding generation? Goes in the qwen hole (qwen embedding)

a screenshot from, "goes in the square hole" tiktok video

000

valarauca @valarauca1.bsky.social · 07/10/2026

Forget the power of technology and science, for so much has been forgotten, never to be re-learned. Forget the promise of progress and understanding, for in the grim darkness of the far future there is only war. There is no peace amongst the stars, only [...] the laughter of thirsting gods.

Fourier transforms below n log n. Showing they're in fact `O((n log n)^1-k)` where k=10^13

010

valarauca @valarauca1.bsky.social · 04/10/2026

I am now realizing why everyone senior in tech is like, "I have no strong opinions about any of this shit", because they've seen it all before multiple times.

000

valarauca @valarauca1.bsky.social · 04/10/2026

1990: make a SOAP API 2010: make a RestFUL API 2030: make an MCP API

100

valarauca @valarauca1.bsky.social · 04/10/2026

4-5x prefill on B70 was literally just refactoring the build file, wtf

000

valarauca @valarauca1.bsky.social · 03/10/2026

Intel has a framework to do p2p GPU communication? Disable it and fallback to vulkan if multiple GPUs are present.

001

valarauca @valarauca1.bsky.social · 03/10/2026

so much shit in llama.cpp intel stuff is unironically written like > Oh no, little johnny decided to run inference on a 15 year old intel iGPU. We should optimize for their experience. Which is actually insane.

200

valarauca @valarauca1.bsky.social · 02/10/2026

lmao (unironically) because the pope took the name "leo" specifically because the last pope leo took a number of very pro-labor stances during the industrial revolution (see: Rerun Novarum 1851).

000

valarauca @valarauca1.bsky.social · 02/10/2026

ad mech gonna canonize this mfer

020

valarauca @valarauca1.bsky.social · 02/10/2026

same thing happened with howard hughes

010

valarauca @valarauca1.bsky.social · 02/10/2026

I once talked to a old school lisp'er in the bay their reason for it never catching on, "nobody could read one-another's macros so every team project went to hell." I think about that a lot.

033

valarauca @valarauca1.bsky.social · 02/10/2026

llvm is **so bad** at splitting vector registers what the fuck? Do you know how many zmm registers we have?!?

000

valarauca @valarauca1.bsky.social · 02/10/2026

DUNE series is extremely horny in a weird creepy way

000

valarauca @valarauca1.bsky.social · 02/10/2026

finally probably embracing the weird sexual elements of the franchise.

110

valarauca @valarauca1.bsky.social · 01/10/2026

Huh, Opus-5.5 just one-shot an extremely esoteric trait algebraic task correctly.

000

valarauca @valarauca1.bsky.social · 01/10/2026

BLAKE3 merkle tree stuff is actually insane. I can take functionally a b3sum of the original `.safetensor` and verify the decompressed data, without losing any speed. So cryptographic verification is free (on an AVX512 processor)

000

valarauca @valarauca1.bsky.social · 01/10/2026

fp8 kind of sucks to compress. bf16 is amazing.

SDXL unet model, data type f16, compressed 0x847, decompress 7.1GB/s
qwen3-vl-8b fp8 compressed 0.796x 4.1GB/s
wan2.2 5B e5m2 + f32, 0.705x & 4.1GB/s
Qwen3.5-27b instruct bf16 0.661x 9.0GB/s

110

valarauca @valarauca1.bsky.social · 01/10/2026

0.65% tensor compression ratio achieved. 0.01bits/weight of the per-tensor order-0 bound. Depending on the local system I'm hitting 10.9GB/s of decompression (per x64 core). Pretty happy with that. credit to the authors: t.co/kYQphr3vNc but I'm not doing GEMM stuff.

t.co

https://arxiv.org/pdf/2606.15789

110

valarauca @valarauca1.bsky.social · 01/10/2026

I'm so sick of fucking tuis.

010

Reposted by valarauca

Brandon Friedman @brandonfriedman.bsky.social · 30/09/2026

Pete is lost in terms of modern war. I use the example of this outgunned Ukrainian army officer who successfully terrorized a Russian unit for weeks because she "did the math." At the time, someone commented, "yeah, in combat, 'I did the math' beats 'I can do 100 pushups' every time."

Ukrainian female soldier: Okay, I said the idea is this: There is enough time to empty one full can of 32 grenades and be gone by the time the return mortar fire reaches us. I did the math.

2349881156

valarauca @valarauca1.bsky.social · 01/10/2026

0.65% tensor compression ratio achieved. 0.01bits/weight of the per-tensor order-0 bound. Depending on the local system I'm hitting 10.9GB/s of decompression (per x64 core). Pretty happy with that.

000

valarauca @valarauca1.bsky.social · 30/09/2026

i know it is shitty woke 1.0 type shit but sometimes when I want to refer to "Americans" in a explicitly derogatory fashion using "Amerikkkans" is undefeated.

121

valarauca @valarauca1.bsky.social · 30/09/2026

Weird Al gets his flowers, but people underrate him. Organizing all the palindromes & making them rhyme is insane.

010

valarauca @valarauca1.bsky.social · 29/09/2026

Love to report non-GAAP figures.

000

valarauca @valarauca1.bsky.social · 28/09/2026

I'm waiting to "learn the downsides" of this sytem, but I've done 2 projects with it so far and it has been rather nice.

010

valarauca @valarauca1.bsky.social · 28/09/2026

turning all the "random bullshit" (linger timeouts in queue, runtime timeout in worker, heart beats, retries, backoff, re-run, cancellation) into configuration & framework makes managing "chains of stateful operations" so much easier.

210

valarauca @valarauca1.bsky.social · 28/09/2026

I'm just gonna start cargo culting temporal because while it isn't super fast (a couple milliseconds between activity/workflow/task hand off) it has so many nice features I can't help but **REALLY** find the strengths out-weighting the cons.

100

valarauca @valarauca1.bsky.social · 28/09/2026

me: here is the VM can you take it from here luna: Nope! I need the following services. And we should setup them up in durable & repeatable way so we can do this again in the future sol/astra: [re-installing all of ubuntu behind your back from pure hallucinated vibes]

020

valarauca @valarauca1.bsky.social · 28/09/2026

gpt-6-luna is just a smol guy who's not really lazy or dub, but seems to give up when stuff goes wrong or assumptions are violated which I'm finding very enjoyable.

120

valarauca @valarauca1.bsky.social · 27/09/2026

At $500/month the ROI on a local setup changes massively, are these people idiots?

000

Reposted by valarauca

Moth! @baphometadata.bsky.social · 26/09/2026

We gotta make hive minds exciting again.

062

valarauca @valarauca1.bsky.social · 24/09/2026

> We're looking for somebody with 3 years of rust-lang experience Me explaining how I stood up a rust-lang application in 2017

010

valarauca @valarauca1.bsky.social · 22/09/2026

> Python helps us move faster Me writing 5KLoc of linter config slop to ensure our system(s) don't nuke themselves.

030

valarauca @valarauca1.bsky.social · 22/09/2026

$The, "Omni-man 'look what they need to mimic a fraction of our power'" meme. Except his face is overlayed with the Rust-Lang logo and the jets are various linters to actually enforce typing within Python.$

1114

valarauca @valarauca1.bsky.social · 21/09/2026

I'm going to get cooked for this, but I'm really happy LLMs exist so I can continue to **not** learn how git works 👍

130

valarauca @valarauca1.bsky.social · 20/09/2026

lowk wild to me as my company "heavy" uses AI but quickly set a policy of, "no matter how you write code, you're still accountable for your contributions", so if you can't defend or explain decisions made by agents, the patches are rejected.

040

valarauca @valarauca1.bsky.social · 20/09/2026

necro politics of labor are literally within capitol vol 2 & 3

082