I run Backpropaganda, Inc, a small garage research lab that tries to work on efficient pretraining and better optimizers.
We initially made money by selling high-speed inference endpoints to individuals who would pay for it: video models for a Hollywood producer, probabilistic models for a casino, an AI scientist harness for a venture debt company.
Today, our work is supported by generous grants from Emergent Ventures and Nous Research.
I come from a family of folk theatre artistes, brought up in the beautiful north-east, at the borders of India. As a theatre artist when I was a child, and as a researcher when I was an adult, I travelled the length and breadth of the country. One picks up a few good stories along the way.
I'm interested in understanding large-scale pretraining (as Dario Amodei says) as a stepping-stone between human evolution and in-vitro learning. Recently, my attention has been on the fact that pretraining produces thickets of task experts that make robust zeroth-order post-training possible. This phenomenon has been noted by a number of works in the past 3 years, but most notably in my opinion by Neural Thickets [Yulu Gan and Philip Isola, 2026] and Finetuning with Only Forward Passes [Malladi et al. 2024].
I am based out of nowhere.