ML research & debugging notes
Achyuthan Sivasankar
M.S. Computer Science at NYU. I write about ML research, debugging stories, and the things I'm still figuring out as I go.
I'm based in New York. I'm especially interested in when neural networks actually learn to generalize — not just whether they do, but what changes inside the model when they finally do.
Right now I'm in Prof. Anna Choromanska's lab, building AD-LiST-JEPA — a self-supervised world model for automotive LiDAR perception — and exploring adaptive computation in world-model architectures. Before that, I spent time on grokking dynamics, expert-collapse benchmarks, graph networks, and a few open-source contributions to the ML tooling I use every day.
This blog is where I write about that work — research notes, debugging stories, and the systems layer underneath it all. For the full picture, myLinkedIn andmy portfolio have the rest.
What I write about
- Mechanistic interpretability and grokking
- World models and self-supervised perception
- MoE routing, expert collapse, and adaptive computation
- Debugging ML frameworks and infrastructure
- Open-source contribution workflows