Tien-Yu Chi

Efficient MLSys and applied AI for science research. Applying for 2027PhD

Now visiting in UCB. Prev @Canva

[Tien-Yu Chi @DianyoChi]

🧬 Can neural networks and GPUs change RNA-seq analysis?

An interactive guide leads to two questions:
🧠 Can a neural network approximate EM?
⚡ Can GPUs make richer alignment practical again?

Bioinformatics feedback welcome:
https://t.co/qPGj1MSmK2

#RNASeq #Bioinformatics

[Tien-Yu Chi @DianyoChi]

@cos_francois @cos_francois Do you have ablation study for different abstraction layer from math definition -> PyTorch -> Triton -> TTIR -> LLVM IR to PTX?

I’m curious for how much abstraction layer does an AI compiler actually need?

[Tien-Yu Chi @DianyoChi]

@SemiAnalysis_ @SemiAnalysis_ may i have the source or code of this info

[Tien-Yu Chi @DianyoChi]

@_albertgu Hey Albert! I’m Tien-Yu, an ML engineer based in Taipei exploring SSM research . Would love to say hi or chat if you’re open to it while you’re in town! Just sent you an email with a bit more context 🙌

[Tien-Yu Chi @DianyoChi]

@anneouyang (2/2) Do you see this approach generalizing beyond PyTorch-to-CUDA translation, for example, toward generating code for other compiler targets like C++ backends or broader compiler abstraction layers (like LLVM compiler)?

[Tien-Yu Chi @DianyoChi]

@anneouyang This is fascinating! reminds me of the “non-precise” kernel generation results from Sakana AI. I have two quick questions:

[Tien-Yu Chi @DianyoChi]

@anneouyang (1/2) Have you considered whether fine-tuning the language model on domain-specific materials like the CUDA Programming Guide or PTX ISA documentation might further improve the quality or correctness of generated kernels?

[Tien-Yu Chi @DianyoChi]

@EyubogluSabri Will using more complex architecture for the Cartridges save training cost from self-study? Just a random and intuition thought, the models might learn more info than current setting 🤔

[Tien-Yu Chi @DianyoChi]

I don’t think people need any high level programming languages in the end. Just assembly languages are enough, high level languages are for human but not for machines.

[Andrej Karpathy @karpathy]

It's 2025 and most content is still written for humans instead of LLMs. 99.9% of attention is about to be LLM attention, not human attention.

E.g. 99% of libraries still have docs that basically render to some pretty .html static pages assuming a human will click through them. In 2025 the docs should be a single your_project.md text file that is intended to go into the context window of an LLM.

Repeat for everything.

[Tien-Yu Chi @DianyoChi]

@SakanaAILabs Why some of the compiled op get much speedup than native one in the table? Shouldn't the compile op already faster than native one?

[Tien-Yu Chi @DianyoChi]

@SakanaAILabs Why some of the compiled op get much speedup than native one in the table? Shouldn't the compile op already faster than native one?

[Tien-Yu Chi @DianyoChi]

@haoailab @haoailab there’s a typo here for FastVideo 😅