Akshay MysoreToronto, 0:00 amPapers

Papers

Research notes and essays on models, agents and the systems around them.

ResearchNovember 2025Scaling the learning rate by progressA small experiment inspired by dopamine: speed up learning when the loss is falling and slow it down when it rises. The overall result is not significant; the late-training gap is suggestive but unreplicated.

ResearchOctober 2025Autograd from scratchA reverse-mode automatic differentiation engine in a few dozen lines of Python, first for scalars and then for matrices, and the one rule that makes all of it work.
ResearchOctober 2025Retrieval as navigation over embedding geometryBuild a graph of a corpus from the shape of its embeddings, then answer a query by walking downhill on that graph. An untested idea, placed against the work it most resembles.
EssayOctober 2025Reading code againLeaning on AI for every change made me worse at the job. The fix was not quitting AI; it was reading every line it wrote and doing the hard parts myself again.
ResearchSeptember 2025JEPA Attach: training a language model on images it never reads aboutA proposal for letting unlabeled images shape a language model's weights through a latent-prediction objective, including why my first version of the idea could not work.
EssayJune 2025What made Sammy 3 workNotes from building a real-time support agent at Sammy Labs: context beat cleverness, memory needed layers, and tracing mattered more than prompts.
EssayJune 2025Building AI products people useStart from a painful workflow instead of a capability, design for failure, and treat speed as part of quality.