I used to love Google Maps Timeline until they removed the web interface. I'm not surprised when they sent a mail saying they accidentally deleted data, I'm actually surprised that there is still a Timeline team.

I'm going to spend most of my time on CUDA and MLX for a while, Apple is sponsoring me to experiment a CUDA backend for MLX. I'll make learnings and make contributions and make money at the same time.

cuda vomiting build errors

This is what I call Japanese tax - the extra money you pay for not speaking Japanese.

I find myself very good at guessing the causes of bugs, but I was almost always wrong when guessing performance bottlenecks.

It is wild that with GPU programing you start measuring performance in units of µs. I wrote some notes on how I tracked down some of my suboptimal code that delayed the kernel execution for 40µs but made whole program 4x slower. github.com/ml-explore/mlx…

NVIDIA should make CUDA runtime's source code available to public, at least under something like proprietary license or guarded access. It is not fun debugging a blackbox.

How does it feel working on VS Code at MSFT, when forks are getting valuated at billions?

My most weird experience with CUDA so far is that, there are so many developers owning PhD, from universities that I could only dream of. I never met one in the first decade of my career.

Once a while I would do a long trip that make myself physically exhausted and mentally refreshed. This is my way for keeping productive.

(just my humble observation) Anthropic seems to hire more engineers while OpenAI hires more researchers — lots of excellent developers I worked with now work at Anthropic — that is probably why they make better products for developers.