I’m a software engineer building high-performance kernels and compilers at Anthropic! Previously at Facebook/Meta (PyTorch, HHVM, ReDex)

Joined December 2022
I’ve seen this “FAANG vibe code” post a few times and tbh this workflow sounds… tedious. “Write a design doc, align stakeholders, iterate on design review, etc” How about: have an awesome idea and build it
1
1
11
806
Claude gives a different response, which seems equally plausible (based on my searching Google for phrases that I can't read 😅). Maybe human translators will still be valuable for disambiguating these, just be more productive.
373
Thinking about AI's and the labor market I wondered about translation as a jobs vulnerable to automation. This industry article claims that machines don't understand cultural nuance and idioms. But of course if you plug their example into ChatGPT, it provides the right idiom
1
2
536
For fun I benchmarked the python and C implementations; to really get the speed of the C version though you have to write the benchmark loop in C too, or else you're just benchmarking the FFI. TL;DR the native version is really fast; it's probably like 10-15 clock ticks.
1
1
2
1,104
I realized you can actual compile native C++ code from Colab and load/run it using ctypes. This is actually how torch.compile compiles code for CPU
1
112
One of the reasons I like hacking in Python is how easy it is to write little helpers to visualize intermediate results. Here's how I visualize the way the grade school long multiplication algorithm achieves the above result:
1
92
To actually convert the byte string "123" to its integer value 123, he multiplies by the constant 0x640a01. I think this is a super clever way to do 2 multiplies and 2 adds in a single instruction.
1
95
`struct` is a cool python module. You can read raw bytes as binary data. So this bit read the byte string b"123\0" as a 4-byte integer:
1
113
What is the setting I need to turn off to stop getting recommended tweets in my notifications? I don’t like getting notifications on posts even from people I follow, much less when it’s right-wing agitprop.
299
Using the allocation timeline, we saw that a bunch of activations were being retained from the end of one backwards pass to the beginning of the next. You can see this in the picture, where there’s a sort of “rainbow” band in the leaky version that’s not in the baseline. 5/
1
42