things i got wrong this year

end-of-year list. not accomplishments, just wrong assumptions. i've been working full-time for about six months.

i thought i'd write more code than i do. most of my time is reading code, reading profile traces, and staring at memory access patterns. the actual code changes are small. a good week might be 50 lines that took 40 hours to figure out. i thought engineering was building. it's more like surgery.

i thought quantization was straightforward. conceptually, sure. cast FP16 to INT8, run the calibration, check the accuracy drop. in practice, every model has layers that quantize badly, and figuring out which ones and why is its own specialty. i spent three weeks on a single attention layer that was producing garbage in INT8 and it turned out the issue was an outlier channel in the key projections that blew up the per-tensor scale. per-channel quantization fixed it. three weeks.

i thought i'd miss school more. i don't. i miss specific things. the Grainger library at 2am. my roommate's climbing sessions. the feeling of learning something purely because it was interesting, without a deadline or a product attached. but i don't miss being a student. working on something real is better. i feel useful in a way i didn't at school.

i thought everyone around me knew what they were doing. they don't. i mean, they're smart and experienced, but there's a lot of "let's try this and see" and "i'm not sure why that works but it does." the codebase has layers of decisions made by people who are no longer here. nobody understands all of it. this was weirdly comforting.

i thought i'd stop thinking about research. the opposite. working on inference made me more curious about what these models are actually doing internally. when you spend all day making a model run faster without understanding what it learns, the question starts to itch.

six months in. still learning. mostly learning what i don't know yet.