Blog
Thoughts on AI research, program management, and building products.
The Bug Was Never in the Weights: Quantizing GLM-5.3-Flash to 4 Bits for Two DGX Sparks
A 320B model that does not fit on my hardware, a week spent squeezing it to 4 bits and measuring what that costs on every one of its 37,152 expert tensors, and a corruption bug that turned out to live in one line of vLLM rather than anywhere in the weights.
The fp8 Mirage: Reverse-Engineering Apple's M4 Max Tensor Path
Apple shipped a tensor matmul in Metal 4.1 and documented almost nothing about how it actually behaves on the M4 Max. So I measured it. Eleven hidden specifications later, the headline is uncomfortable: the fp8 path is emulated, not accelerated — and that turns out to be the most useful thing you can know before you write a single kernel.
What If LLMs Had Life? The Strange, Sobering Lessons of UYIR
I spent six months trying to make a language model evolve — to give it a population of adapters that compete, die, reproduce, and pass knowledge to their neighbors. The data did not love the story I wanted to tell. This is what it told me instead, why I am writing it down anyway, and where I want to take this next.