Matt Krzus

research enjoyoor. jiu jitsu black belt. roadhouse is the greatest movie ever made.

[Matt Krzus @mattkrzus]

skip diffusion boogaloo round 900: how much diffusion do i _really_ need: tail forgives noise/compounds mistakes, one real step beats 1000x more training data, jump distance is free. yeehaw.

teacher vs. naive vs. prefix jumping:

[Matt Krzus @mattkrzus]

less is the lord baron of kernels

[Less Wright @lessw2020]

Blog post showcasing some of my kernel work at Cursor:
https://t.co/wUWXdXYzaM

[Matt Krzus @mattkrzus]

the researcher who also cares enough to be a kernel warlord is everyone’s friend.

[Horace He @cHHillee]

Presenting my grand unified theory of ML researcher impact: Your impact is directly proportional to how much pain you cause to infra.

Fundamentally, you can only inflict pain upon infra if your approach actually works. And the better your approach works the more pain infra is forced to endure.

So, to give some examples:
- MoE's add a ton of data-dependent computation => pain (shazeer++)
- GDN/KDA are the most complex architecture I've been forced to care about and a very annoying matrix inversion => pain (sonta++)
- Muon is much more annoying than Adam and causes annoying restrictions on parallelism => pain (keller/jeremy++)
- RL scaling forced many researchers to care about LLM inference and RL infra as a category => pain (tworek++)

Even papers like Attention Is All You Need have lead to significant pain! Before transformers were invented everyone was running small jobs and I never needed to think about kv-caches or 6D parallelism.

[Matt Krzus @mattkrzus]

the radfords of the world have in abundance a quality that’s hard to distill to a single word. taste? creativity? something ethereal that complete mechanical knowledge alone doesn’t convey.

[Andrew Ho @andrewho03]

This is untrue. I am a mediocre researcher and even despite access to internal models and thousands of GPUs, I would not (at this moment) be able to accomplish something equivalent in significance to what Alec Radford could do in a single week several years ago.

[Matt Krzus @mattkrzus]

sidakov has the most teachable system of all time. just so good.

[Wrestling Rabbit Hole @wrestlingrh]

Sidakov's Handfight System
https://t.co/LaYS0Qf5nI

[Matt Krzus @mattkrzus]

i hate the gi (just kidding bb i love you)

[Matt Krzus @mattkrzus]

come hang - i’ll be giving free leglocks

[uRun @urunml]

https://t.co/kbL1maA4T8 launch party - Wednesday, April 29 · 6PM:

🕹️ Arcade games
🍹 Open bar
💻 Live demos
🥽 Meta Quest Giveaway

Spots are limited - click the link to grab your invite.

👉 https://t.co/Hx5owyyoMX

[Matt Krzus @mattkrzus]

this is why i have a black belt. when my job is completely automated, at least i can fallback on teaching people how to enter cross ashi from seated positions.

[himanshu @himanshustwts]

a significant % of ml researchers might be hooked by what happened in ONE day.

ai seems to be doing a research loop fascinatingly well (understand the problem + propose a change + train/test it + measure results + keep the better version + repeat) and genuinely reducing research friction.

we are early to automated experimentation, frontier scale could be an interesting watch.

[Matt Krzus @mattkrzus]

blog: https://t.co/56MVAgxw66

code: https://t.co/0blnNqPTIz

[Matt Krzus @mattkrzus]

as i try to solve the problem of making a robot lawyer go brrr, i found myself on a small grpo sidequest

tldr: its probably a nothing burger, but i hacked together a faster than vanilla grpo implementation w triton/cuda streams. it was beating torch compile, so thats cool

[Matt Krzus @mattkrzus]

link to lib: https://t.co/edOiALDYDo

[Matt Krzus @mattkrzus]

based

i had been grpo-ing over reasoning tasks from ace attorney

need more legal games like something fierce

[Hao AI Lab @haoailab]

When Ilya Sutskever once explained why next-word prediction leads to intelligence, he made a metaphor: if you can piece together the clues and deduce the criminal’s name on the last page, you have a real understanding of the story. 🕵️‍♂️

Inspired by that idea, we turned to Ace Attorney to test AI's reasoning. It’s the perfect stage: the AI plays as a detective to collect clues, expose contradictions, and uncover the truth.

We put the latest top AI models—GPT-4.1, Gemini 2.5 Pro, Llama-4 Maverick, and more—to the test in Ace Attorney, to see if they could shout Objection! ⚖️, turn the case around, and uncover the truth behind the lies.

[Matt Krzus @mattkrzus]

https://t.co/urdSl2LrYQ

[Matt Krzus @mattkrzus]

mech interp has been growing on me as of late so i wrote some code/words on replicating the circuit kings

[DeepSeek @deepseek_ai]

🚀 Day 5 of #OpenSourceWeek: 3FS, Thruster for All DeepSeek Data Access

Fire-Flyer File System (3FS) - a parallel file system that utilizes the full bandwidth of modern SSDs and RDMA networks.

⚡ 6.6 TiB/s aggregate read throughput in a 180-node cluster
⚡ 3.66 TiB/min throughput on GraySort benchmark in a 25-node cluster
⚡ 40+ GiB/s peak throughput per client node for KVCache lookup
🧬 Disaggregated architecture with strong consistency semantics
✅ Training data preprocessing, dataset loading, checkpoint saving/reloading, embedding vector search & KVCache lookups for inference in V3/R1

📥 3FS → https://t.co/JRZ2eYk1aK
⛲ Smallpond - data processing framework on 3FS → https://t.co/BoUA6YNjZK

[Matt Krzus @mattkrzus]

the race to give your legal model the actual personality of a real life judge, with that judge’s morals/ethics/etc is on.

c thomas, posner, scalia, etc.

[Nathan Lambert @natolambert]

In 2023 and 2024 labs perfected the listicle with post-training/rlhf.
In 2025 the personality training of models is on center stage.

There's almost 0 academic work on Character Training and almost 0 work on the web writ large.

We need to change that - it starts with this post.

[Matt Krzus @mattkrzus]

@Ishaank1999 they clappin

[Matt Krzus @mattkrzus]

based whale.

i’ve got like 20 tweets and half are whale quote tweets.

[DeepSeek @deepseek_ai]

🚀 Day 4 of #OpenSourceWeek: Optimized Parallelism Strategies

✅ DualPipe - a bidirectional pipeline parallelism algorithm for computation-communication overlap in V3/R1 training.
🔗 https://t.co/NR3047Nnpo

✅ EPLB - an expert-parallel load balancer for V3/R1.
🔗 https://t.co/IilrQUfn49

📊 Analyze computation-communication overlap in V3/R1.
🔗 https://t.co/tY1WujA1Lw

[Matt Krzus @mattkrzus]

xmas in jan

[lujianqiao @lujianqiao3]

🚀 Exciting News! We’ve just released [Native Sparse Attention Triton](https://t.co/CIBlMh6gBG) – a full Triton-based implementation of NSA (fwd + bwd) with guaranteed efficiency! 🚀

[Matt Krzus @mattkrzus]

bunch of beauties

[DeepSeek @deepseek_ai]

🚀 Day 1 of #OpenSourceWeek: FlashMLA

Honored to share FlashMLA - our efficient MLA decoding kernel for Hopper GPUs, optimized for variable-length sequences and now in production.

✅ BF16 support
✅ Paged KV cache (block size 64)
⚡ 3000 GB/s memory-bound & 580 TFLOPS compute-bound on H800

🔗 Explore on GitHub: https://t.co/4JvJTn5HX2