about

i'm Shyam Pathak, an ai/ml engineer and researcher. this site is a long-running notebook — essays on whatever i'm chewing on, terse summaries of papers i've read carefully, and write-ups of small projects when they're worth sharing.

most of what shows up here lives at the intersection of training, evaluation, and interpretability. i'm particularly drawn to mechanistic-interp tooling, scaling recipes, and the gap between what benchmarks measure and what models actually do.

nothing here is ai-generated unless explicitly labelled. nothing here speaks for any employer.

elsewhere

colophon

built in astro with tailwind, mdx, and pagefind. served from github pages. typeset in fraunces (display), newsreader (body), jetbrains mono (mono). dark mode is the default — toggle in the header. forest greens in dark, parchment creams in light. no analytics, no comments, no third-party scripts.

◇ if you spot a mistake, tell me — i'll fix it.