Tommy Han
Tommy is a software engineer in Hong Kong who loves to work on open source projects in different areas, including toolchains, compilers, apps, and IDE development.
Sessions
Large Language Models may start in frameworks like PyTorch, but getting them to run efficiently on CPUs, GPUs, and specialized accelerators increasingly depends on compiler infrastructure.
As machine learning systems have grown more complex and hardware has become more specialized, the compiler layer has become a key bridge between model code and machine execution.
This talk introduces AI compilers from a practical perspective, covering why they matter, what kinds of problems they solve, and how open-source projects such as LLM MLIR and OpenXLA/XLA fit into the broader stack.
It will also introduce a bit about tinygrad, which is trying to take another approach rather than LLVM, to compiler models running on the hardware..
It also explains how ideas such as intermediate representations, lowering, and backend targeting connect frameworks like PyTorch to real hardware.
A small demo based on PyTorch and various compilers (e.g. LLVM and tinygrad) helps make the overall pipeline more concrete and easier to understand.
AI tools are making it easier than ever to generate code, explore ideas, and open pull requests quickly. But faster output does not always lead to better PRs. In practice, AI-generated changes often come with familiar problems: larger diffs, weaker context, less clear intent, and higher review overhead.
In this talk, I will share a practical view of where AI genuinely helps in day-to-day development, where it starts to create problems, and why I still choose to manually craft most pull requests. Rather than arguing against AI, this session focuses on a more useful question for engineers and open-source contributors: how do we use AI as a tool without giving up ownership, clarity, and reviewer empathy?
I will also briefly use my own open-source contribution experience, including work related to Zed, as an example to show why a reviewable, well-structured change is often more valuable than a quickly generated one.
The goal of this talk is to provide practical suggestions about how to build a proper PR to contribute to open-source in a healthy way, and understand why keep context understanding personally is important.