krishna

independent ai researcher trying to contribute my share in the inevitable machine god.

probehover the figure. every character is a ray; this line reports what happened to it.fig. 1schwarzschild black hole, thin disc 3–11 rs, i = 9°. one null geodesic integrated per character (0×0), doppler + gravitational shift. hover to probe a ray, drag to tilt, double-click to reset.
§1 nowhacking stuff around, reading papers and implementing them.§2 shippedcubeeval — can LLMs rotate shape?mental-rotation benchmark for language models: a shape, a rotation, a question about the result.results ↗fig. 2 R = R_y(35°)·R_x(28°) · drag to rotatesidesignull ↗ — chrome extension that uses jev to filter x + youtube down to what you actually want to see.all projects →§3 recent2026-09-28AI-ML2026-09-28GradMem2026-09-28RL's razor2026-09-28SPD via capability selective subspace projection§4 beforeai stuff at composio. independent contract work at startups, integrating ai agents.§5 notes[1]GradMemai-ml2026-09-28[2]RL's razorai-ml2026-09-28[3]SPD via capability selective subspace projectionai-ml2026-09-28[4]deep-seek v3.2 sparse attentionai-ml2026-09-28[5]genieai-ml2026-09-28[6]locating and editing factual associations in GPTai-ml2026-09-28[7]memory R1ai-ml2026-09-28[8]weight-sparse transformers have interpretable circuitsai-ml2026-09-28all notes →§6 elsewheregithubtwittersource
set in ibm plex mono · rendered in characters0 fps · 0 rays