Shokhrukh Mir (@shmonchik) header
Shokhrukh Mir (@shmonchik) profile picture

this bio is a placeholder for a career i haven't invented yet. [email protected]

Joined September 2022
Photo by Shokhrukh Mir (@shmonchik)Photo by Shokhrukh Mir (@shmonchik)Photo by Shokhrukh Mir (@shmonchik)Photo by Shokhrukh Mir (@shmonchik)Photo by Shokhrukh Mir (@shmonchik)Photo by Shokhrukh Mir (@shmonchik)Photo by Shokhrukh Mir (@shmonchik)Photo by Shokhrukh Mir (@shmonchik)Photo by Shokhrukh Mir (@shmonchik)Photo by Shokhrukh Mir (@shmonchik)
I’m staying one step ahead.
If you still have your tech stack in your bio, you're probably gonna be unemployed in 6 months
20
Shokhrukh Mir retweeted
Welcome back to Burger Mart! Please enjoy a hot BM along with an update on Season 5 of...
956
14,781
2,249
108,421
11,186,054
Most AI integrations bypass core security boundaries. I built a domain-agnostic modular monolith in Go where AI is treated as just another actor - bound by the exact same RBAC, tenant isolation, and audit logs as human users. How the architecture is structured: • Thin Core: Identity, RBAC, multi-tenancy via PostgreSQL RLS, and a Commands → Events → Audit pipeline. • Governed AI: A separate gateway handling keys, rate limits, and per-tenant token budgets before requests ever reach the core. • Plugin Domains: Business modules sit strictly on top (built a university IS first: grades, tasks, and a double-entry ledger). • Contract-First: OpenAPI as the single source of truth, sqlc, transactional outbox, and full observability. Where do you draw the line between DB-level isolation (Postgres RLS) and application-layer checks in modular monoliths? #golang #softwarearchitecture
3
2
6
279
For now in Russian, but soon I will translate if necessary
14
My first photoshoot for an article (the photographer was in a rush)
1
46
Nice progress on qwen-vx - the gap to llama.cpp is getting really small. I've been working on NNL (Neural Network Lego), a modular architecture for composing specialized blocks over a shared bus with strict isolation on the first pass and causal refinement on the second. The goal is a system that doesn't lie and doesn't cascade errors (hot-pluggable blocks, selective "I don't know", conflict detection). Currently pure Python/numpy, domain-tested on a medical/legal classification task. Might be interesting as a higher-level composition layer on top of something like Vx kernels, or as a use-case for the type system (placement + isolation guarantees). Repo: github.com/smworklair/NNL
qwen-vx closing the gap with llama.cpp qwen-vx is a small inference program for Qwen coder models, built on the GPU kernels of Vx's standard library. Dense Qwen3-Coder-30B-A3B and the mixture-of-experts model in 4-bit Q4_0 (fits in 32GB Macbook M4). Some small kernels remains to be fused.
4
121
An instance of "AI racism" occurred. I told a Chinese person that Claude is better than DeepSeek, GLM, and Kimi. We argued for half an hour, and the debate ended with us agreeing that Russia, for its part, has SberbankGPT and YandexGPT.
1
3
81
Rewriting the new NEX CLI in C x C++ x Zig. Doing the whole thing in Python was a mistake. Wrong language for a compiler and the dependency was just slowing everything down. Level 1 crash tests (e2k VLIW): native code is already in the tens of microseconds per block. Python was still sitting at ~0.6ms best case. Still only level 1. Once it is stable I am moving to level 2. Too early to talk real efficiency numbers but the stack is already feeling solid. @adityazero_ @ayaka14732 @nanoalpaca #compilers #zig #cpp #systemsprogramming #vliw
1
4
265
Shokhrukh Mir retweeted
I have a counterexample on TPU v7x. In a matmul kernel, issuing matmuls in bursts triggers power-management throttling, which in turn stalls weight delivery. Spacing out the matmuls yields a 25.9% speedup over an issue-when-ready greedy scheduler and 26.4% over vanilla Pallas.
Anyone who worked with wide instruction bundles / static scheduling on TPU, Hexagon, or AMD AI Engine — curious how close a greedy scheduler gets to optimal on your machines. I measured this on Elbrus (classic VLIW). Greedy is already optimal in 94% of cases, 0.51% cycle loss. Small neural net recovers most of the rest and never makes it worse. CLI that shows greedy vs exact optimum vs model on any .s file: github.com/smworklair/NEX-CL… @PatrickToulme @NormJouppi @adityazero_ @AminVahdat @JeffDean @Qualcomm @QCOMResearch @AMDembedded
11
15
4
241
15,158
Shokhrukh Mir retweeted
Replying to @shmonchik
You can look at our (Sergei and I) presentation from 12 years ago. llvm.org/devmtg/2014-10/Slid… Greedy is optimal on ML workloads because the variety of workload is small (so far but might change in the future). Packet density is usually < 3 even with instruction scheduling. If you really want to step up the bundle packer game then borrowing ideas from GVNHoist.cpp (from LLVM repo) might help. If you are generally interested in VLIW hiraditya.github.io/posts/wh… might be interesting. PS: Never thought Jeff Dean and my name will appear in the same sentence, You made my day. Thank you! 🙏
2
1
9
463
Shokhrukh Mir retweeted
Very few among us know the joy of working on the TPU VLIW bundle packer.
Anyone who worked with wide instruction bundles / static scheduling on TPU, Hexagon, or AMD AI Engine — curious how close a greedy scheduler gets to optimal on your machines. I measured this on Elbrus (classic VLIW). Greedy is already optimal in 94% of cases, 0.51% cycle loss. Small neural net recovers most of the rest and never makes it worse. CLI that shows greedy vs exact optimum vs model on any .s file: github.com/smworklair/NEX-CL… @PatrickToulme @NormJouppi @adityazero_ @AminVahdat @JeffDean @Qualcomm @QCOMResearch @AMDembedded
2
4
60
4,349
Shokhrukh Mir retweeted
Replying to @shmonchik
Yes i have worked on the bundle packer before. This is cool!
1
1
8
678
Anyone who worked with wide instruction bundles / static scheduling on TPU, Hexagon, or AMD AI Engine — curious how close a greedy scheduler gets to optimal on your machines. I measured this on Elbrus (classic VLIW). Greedy is already optimal in 94% of cases, 0.51% cycle loss. Small neural net recovers most of the rest and never makes it worse. CLI that shows greedy vs exact optimum vs model on any .s file: github.com/smworklair/NEX-CL… @PatrickToulme @NormJouppi @adityazero_ @AminVahdat @JeffDean @Qualcomm @QCOMResearch @AMDembedded
Жадный планировщик на Эльбрусе уже оптимален в 94% случаев. Потери всего 0,51% тактов. Маленькая нейросеть возвращает большую часть оставшихся потерь и никогда не делает хуже жадного. Сделал CLI, который это показывает на вашем .s-файле: жадный, точный оптимум и модель рядом. Репозиторий: github.com/smworklair/NEX-CL… Кто работал с широкими командами - насколько это похоже на то, что вы видите?
3
2
2
45
16,948
Shokhrukh Mir retweeted
20,163
485,069
42,610
1,366,636
Жадный планировщик на Эльбрусе уже оптимален в 94% случаев. Потери всего 0,51% тактов. Маленькая нейросеть возвращает большую часть оставшихся потерь и никогда не делает хуже жадного. Сделал CLI, который это показывает на вашем .s-файле: жадный, точный оптимум и модель рядом. Репозиторий: github.com/smworklair/NEX-CL… Кто работал с широкими командами - насколько это похоже на то, что вы видите?
1
1
8
2,200
gemini heroine 4 I stepped out, and now there’s some finger here.
1
5
150
There are more errors in this post than there are letters in the English language. Take a look at my GitHub—I’m not great at self-promotion, but please consider me for an internship; I’m genuinely smart. github.com/smworklair
3
62
Shokhrukh Mir retweeted
Actual footage of the mechanics behind WinRAR compression! 🤣😂🤣
WinRAR vs a 67 GB File 🥀
76
319
11
4,084
69,840