Jędrzej Maczan

ML research / open science @Cohere_Labs / writing @pagedout_zine / contributor @PyTorch / tiny-vllm
cover art

[Jędrzej Maczan @jedmaczan]

I built a tiny-vllm in C++ and CUDA
- paged attention
- continuous batching
- educational
- 100% human-written™

And now I writing a course where you will build your own vLLM yourself. Still work in progress, I'll finish by the end of April. All for free ofc, just a GitHub repo

[Jędrzej Maczan @jedmaczan]

that's a weird flex, why you don't allow people finally close this info bar?

EDIT: F12 (devtools) -> console -> paste:

document.querySelector('div[role="status"][aria-live="polite"][aria-atomic="true"]')?.remove();

[Sara Hooker @sarahookr]

Today I learnt how Albert Einstein treated his first wife.

And I can’t unsee it. Whattt how does no one mention this.

[Jędrzej Maczan @jedmaczan]

@yacinelearning Thanks, I will give it a shot too

[Jeff Dean @JeffDean]

@uwcse @MBalazinska Here are a few of the pieces of advice that I shared:

[Mathematica @mathemetica]

Hong Wang (MIT PhD 2019, now Associate Professor at NYU Courant & Permanent Professor at IHES) together with Joshua Zahl has achieved a major breakthrough in geometric measure theory.

In February 2025, they posted a 127-page preprint proving the long-standing three-dimensional Kakeya set conjecture.

A Kakeya set in ℝ³ is a compact set that contains a unit line segment in every possible direction. For over a century, mathematicians wondered how “small” such a set could be. Wang and Zahl proved that these sets must have full Minkowski and Hausdorff dimension 3; they cannot hide in lower-dimensional space, even though they can have zero volume.

This result settles a fundamental question with deep connections to Fourier analysis, the restriction problem, PDEs, and incidence geometry. Their work builds on years of incremental progress (including their own earlier papers from 2022 onward) and introduces powerful new volume estimates for unions of tubes and convex sets.

[Fei-Fei Li @drfeifei]

Scientific research is fundamental to advancing civilization and helping people globally to solve the most critical problems, from medicine to materials, from brain science to physics, and much beyond. This is only possible when scientists have access to the best tools of the time to conduct scientific research, including having access to AI-based tools.

[Michał Moch @MichalMoch87]

Reprezentacja Haiti na swoich koszulkach umieściła. POLSKĄ FLAGĘ! 🇵🇱

To nie przypadek ani pomyłka a hołd dla naszych przodków, żołnierzy Legionów Polskich, którzy wysłani przez Napoleona nie walczyli z niewolnikami, stanęli po ich stronie i pomogli Haiti wywalczyć niepodległość.

[Weixin Wu @celia250527]

Even as an insider, our MAI-Thinking-1 report taught me a ton.
We’re sharing the exact mechanics of building a frontier LLM from scratch w/o distillation: data, backbone, RL, infra...
This radical transparency is our DNA at MAI. Eager to learn? Join us 🚀
https://t.co/yetH7KgeKK

[Jędrzej Maczan @jedmaczan]

"They Won't Say Anything" Anita Maczan
150.00x120.00 cm
2026
Acrylic, oil, and oil pastels on canvas

[Jędrzej Maczan @jedmaczan]

@eriklocalhost Really nice project. Posted it on HN for you and it’s on the front page right now https://t.co/6Lcxf6CLhj

[Jędrzej Maczan @jedmaczan]

online gallery: https://t.co/3ljSYrXsTC
ig: https://t.co/RNTi9vZFdY

[Jędrzej Maczan @jedmaczan]

"They Won't Say Anything" Anita Maczan
150.00x120.00 cm
2026
Acrylic, oil, and oil pastels on canvas

[Jędrzej Maczan @jedmaczan]

@MillionInt Faith makes CUDA

[Jędrzej Maczan @jedmaczan]

the Internet people, do you like posts/threads with math proofs? I learn more math recently and I would be happy to share things if anyone finds that interesting too

[Jędrzej Maczan @jedmaczan]

From Boltzmann-Gibbs distribution to numerically stable online softmax

[1/13]

[Jędrzej Maczan @jedmaczan]

Based on:
1) my recent article "From Boltzmann-Gibbs distribution to numerically stable online softmax on GPU with proofs"
https://t.co/rsLeUA96mc

2) "Online Softmax" talk in Cohere Labs Open Science Community - ML Math
https://t.co/2h4Sx5dVsO

[13/13]

[Jędrzej Maczan @jedmaczan]

From Boltzmann-Gibbs distribution to numerically stable online softmax

[1/13]

[Jędrzej Maczan @jedmaczan]

References (4/4):
[4] Zihao Ye. From Online Softmax to FlashAttention. 2023. url: https://t.co/w9Oh1VkepH

[12/13]