Sudarsh
I work on reasoning systems, inference optimization, and multimodal AI. My world model is simple: build systems that expose how models think, not just what they output.
- experiment One neuron is all you need
- kernel A LayerNorm kernel in CUDA, worklog
- math Quasi-RH hill-climb
Selected work
- Experiment One neuron is all you need Can one MLP neuron make Qwen3-8B think it is God? Single-neuron steering, replicated on a 6 GB GPU with llama.cpp control vectors.
- Experiment Dust on GPT-2 A small probe of whether Dust's zeroth-order update finds descent that backprop misses, on GPT-2 124M.
- Contribution Quasi-RH hill-climb Pushing the conditional Quasi-RH zero-free half-plane past 7/8, with exact certificates and audits.
- Contribution Integer multiplication bounds Conditional integer multiplication with κ > 2⁻¹⁵: community proofs, exact certificates and reproducible audits.
Recent writing
- 2026-09-06 How They Stole My Braj From Me An essay on the difference between the geographic Braj and the ideal it was supposed to embody — on how teachers, ritual, and darshan gave way to commodification, inherited authority, and spiritual theater.
- 2025-02-17 Optimizing a Layer Normalization Kernel with CUDA: a Worklog An iterative guide to writing and optimizing a CUDA layer normalization kernel — from a naive single-thread implementation to vectorized loads — benchmarked against PyTorch.