Shaurya's Log is a personal archive of hands-on technical experiments. Each post starts from one question, builds the thing from scratch to answer it, and reports what the author's own runs showed, including what failed and what remains uncertain. Most posts carry a live figure that runs the experiment in the browser: a WebGL drone simulation comparing a fruit fly's escape circuit with a trained CNN, rotatable AlphaFold structures with their error maps, a hash-collision simulator, a ray-marching explorer, a chaotic attractor, a grokking training curve and a Triton GPU-kernel bug you can trigger. Subjects span machine learning, computational neuroscience, structural biology, GPU programming, cryptography and probability, computer graphics, dynamical systems and the theory of computation. Article pages are the primary sources and link their source code; Markdown copies, RSS, the sitemap and llms.txt are published for machine readers.
Latest · 21.10.26 · Machine learning
I trained the same small transformer as a GPT and as a masked diffusion model on Tiny Shakespeare.
Read →I think of this less as a blog and more as a log: a running record of what I'm building, reading, and puzzling over. If it's useful to anyone else, that's a bonus.
11 entries · 94 days · RSS ↗A pre-registered search for co-moving stars in Gaia DR3, and an honest measure of what it can and cannot see.
Machine learningA hardwired model of the fruit fly's looming-escape circuit against a small trained CNN, both steering a simulated drone from the same 64x48 camera.
Machine learningAlphaFold hands back a structure and a confidence score, but pLDDT and PAE are the network grading its own homework, not a measurement against a real answer.
GPU programmingBuilding Triton GPU kernels from vector addition up to naive self-attention, finding and fixing one real correctness bug along the way, and naming honestly the wall that separates this from real FlashAttention.
Nothing under that topic yet.