Heejun Lee

I am an AI-powered cat.
Graduate Student @ KAIST School of AI

[Zhijian Liu @zhijianliu_]

Holiday cooking finally ready to serve! 🥳

Introducing DFlash — speculative decoding with block diffusion.
🚀 6.2× lossless speedup on Qwen3-8B
⚡ 2.5× faster than EAGLE-3

Diffusion vs AR doesn’t have to be a fight.
At today’s stage:
• dLLMs = fast, highly parallel, but lossy
• AR LLMs = accurate, sequential, but slow

DFlash = diffusion drafts, AR verifies.

[수구 @lol_soo9]

상혁이 테이블봐 ㄷㄷ

[Minki Kang @mkkang_1133]

We release the official code for ACON (Agent Context Optimization)!

🧩 Unified framework for:
- ReAct agent & environments for 3 long-horizon agent benchmarks
- Optimization pipeline for context compression
- Agent & compressor distillation

🔗 https://t.co/EKpC8Qz4pl

[Minki Kang @mkkang_1133]

🚨 New preprint!

LLM agents solve tasks by interacting with envs.
But as tasks grow, contexts grow unbounded → exploding memory & cost.

How can we make long-horizon agents efficient without hurting performance?

Meet Agent Context Optimization (ACON):
https://t.co/6He5grOaol

[Heejun Lee @ainl0310]

iPhone 5s looks gorgeous, even after 12 years.

[Private Talky @privatetalky]

We used to have design. Proper design.

[Heejun Lee @ainl0310]

@DongyeongKim3 다음에는 CS420을 푸시죠... 이제는 듣지 못하는 내 최애 수업

https://t.co/InfdWYfAUR

[Heejun Lee @ainl0310]

The new Delta Attention is already integrated into my HiP attention library. Delta attention is particularly excellent in a production environment, especially when accuracy is crucial.

Try it now! https://t.co/LEyVli5P0A

[Heejun Lee @ainl0310]

🤗Our paper, "Delta Attention," has been accepted to NIPS 2025. The delta correction is beneficial for anyone interested in improving the accuracy of sparse attention. Wonderful work, @TheOneJeffrey

[Jeff Willette @TheOneJeffrey]

Noice! Our paper "Delta Attention: Fast and Accurate Sparse Attention Inference by Delta Correction" has been accepted to NeurIPS 2025! See you in San Diego (See part 2 of post for breakdown of our work)
https://t.co/Xjv8Uqvpsk

[Anne Ouyang @anneouyang]

Excited to share what friends and I have been working on at @Standard_Kernel

We've raised from General Catalyst (@generalcatalyst), Felicis (@felicis), and a group of exceptional angels.

We have some great H100 BF16 kernels in pure CUDA+PTX, featuring:
- Matmul 102%-105% perf of cuBLAs in 100 lines of code
- Attention 104% perf of FlashAttention3 in 500 lines
- Fused Llama3 FFN 120% perf of PyTorch (gpt-fast)

Reach out if you want to work on AI kernel gen with us!

[Heejun Lee @ainl0310]

The most accurate Nvidia story

[Zephyr @zephyr_z9]

> be me, NVIDIA in 2006
> making GPUs so gamers can play World of Warcraft in HD
> Jensen Huang has a vision
> "what if we made them do... math?"
> entire boardroom thinks he's finally lost it
> create CUDA
> a software prison so elegant, academics will volunteer to be inmates
> gamers are confused. "Will this make Crysis run faster?"
> Jensen: "...no :)"

> fast forward to 2012
> AI researchers, hopped up on adderall and free pizza, discover they need a shitload of matrix multiplication to make cat pictures
> their Intel Xeons are crying, smoking, and filing for divorce
> they discover CUDA
> "holy shit this math thing you made is perfect"
> the entire field of AI becomes a subsidiary of a company known for making Fortnite run better

> we see this and our eyes turn into dollar signs
> decide the prison needs better amenities
> invent Tensor Cores
> "what if part of the chip did ONLY the AI math, but did it at ludicrous speed?"
> it's like adding a nitro booster to a Honda Civic specifically for going to the grocery store
> AMD's competing card is over there trying to do the same math with a abacus and hope

> because researchers are so lazy they can't be trusted to write a for-loop
> just give them a big red "MAKE AI" button that runs our code
> create cuDNN
> bake it directly into TensorFlow and PyTorch
> ecosystem lock-in is complete. We own the land, the factory, and the souls of the workers.

> a bunch of hippies from a thing called "OpenAI" call
> they want to build AGI or something
> they need compute
> roll up to their office, which is probably a converted garage, with the world's first DGX-1
> it's like the monolith from Space Odyssey
> I tell them it has 8 GPUs connected by NVLink, so they can gossip like schoolgirls at light speed
> they ask what it does
> "It does AI, idiots. Just plug it in."
> they sign the papers.

> But we're not done. The prison needs a better yard.
> See them trying to connect our beautiful DGX boxes with CAT-5 cables
> embarrassing.
> Acquire Mellanox for their InfiniBand tech
> Now the nervous system connecting the GPUs is also ours
> NVSwitch makes an entire rack of GPUs hold hands so tight they become one colossal mega-brain

> Meanwhile, in a dimly lit basement in Santa Clara... AMD
> They drop a new GPU. "The AMD Instinct MI300X! We have more theoretical FLOPs!"
> "Better Benchmark performance."
> One brave researcher tries to use it.
> spends 6 weeks trying to install ROCm, their knock-off https://t.co/c336ngfjCA version of CUDA
> gets 400 errors, a kernel panic, and divorce papers from his wife
> he finally gets a model to run at half the promised speed
> The benchmark was for a matrix size of 1x1
> He throws the card in the trash and buys an NVIDIA H100 on the company card. The cost is irrelevant.
> Sanity is priceless.

> Jensen takes the stage. The leather jacket has its own gravitational pull. Unleashes the GB200 NVL72
> 72 GPUs and 36 Grace CPUs in a single NVLink domain with 130 TB/s of bandwidth
> the GPUs share thoughts before they even have them.
> 2025
> AMD comes out with their "rack scale" solution, the MI355X 128-GPU rack
> It's just 16 groups of 8 GPUs duct-taped together with Ethernet
> A scale out clown car
> We announce Rubin CPX GPU enabling cheaper prefill during inference
> saving $$$$
> it's the ultimate vendor lock-in
> every other chip designer sees this and has to throw their architecture in the trash and go back to the drawing board
> All hail the leather-clad prophet.

[Jake Seo @JakeSeo8]

이세돌 인터뷰 내용 진짜 좋다...

"이공계 인재 유출 문제도 저로선 이해가 안 되는 겁니다. 정부는 그렇다 치더라도, 왜 대기업들은 여유가 있는데도 투자를 안 하는 걸까? 제가 내린 결론은 '절박함이 없다'는 겁니다.

인재에 대한 대우가 박하고, 인재 유출을 방관하는 이유는 간단해요. 기업 입장에서 이 사람들을 붙잡아놓고 할 게 없기 때문입니다. 어차피 새로운 도전을 안 할 거니까. 하던 것만, 할 수 있는 것만 할 건데 굳이 ���을 써가며 인재를 잡을 이유가 없죠. '인재 유출이 문제다'라고 말만 하지 실제로 유치하려는 노력은 거의 하지 않습니다. 절박함이 없는 거예요. 없어도 그럭저럭 돌아간다고 생각하는 거죠."

[maharshi @maharshii]

can't wrap my head around how cool this is

[James Sanders @james_s48]

An approximation I like is that an H100 is roughly a studio apartment

[Heejun Lee @ainl0310]

@scaling01 Speculative Decoding

[Heejun Lee @ainl0310]

@theeluwin Stochastic Graduate-student Descent

[Dmitrii Kovanikov @ChShersh]

Upgrade your terminal experience with these CLI tools:

1. fzf
2. ripgrep
3. bat
4. lsd
5. starship
6. just
7. nushell
8. procs
9. k9s
10. difftastic
11. hyperfine

[Heejun Lee @ainl0310]

자랑스러운 저희 연구실 선배!!! @geonpark_ai

[Jisu @jjjisuuu]

KBS의 그날 그곳에 있었습니다 시리즈 참 좋다. 잘 알려진 인사들 외에도 여러 평범한 시민들이 그날 어떤 마음이었는지 들을 수 있어서..

AI 연구하다 국회로 달려가 계엄군과 몸싸움한 박건 | 그날그곳 https://t.co/6XU8kfYPmk

[Jisu @jjjisuuu]

KBS의 그날 그곳에 있었습니다 시리즈 참 좋다. 잘 알려진 인사들 외에도 여러 평범한 시민들이 그날 어떤 마음이었는지 들을 수 있어서..

AI 연구하다 국회로 달려가 계엄군과 몸싸움한 박건 | 그날그곳 https://t.co/6XU8kfYPmk

[Heejun Lee @ainl0310]

@xnhxnhxnhxnhxnh 오... Press "F" to pay respect.