Manifesto

For notes on civilization, my views on the \(21^{st}\) Century, and how it is time we began to think seriously about the eventual colonization of the sun, see: the demotheosis manifesto. If instead (for whatever reason), you are here to know about me, my Life, Universe and Everything, you’re already parked at the right place.

My Background, without the Granularity

My current work always redirects from here.

Current interests are primarily in the realms of distributed computing and optimization theory¹. I have in the past, in various capacities, worked on: pretraining large, production-grade transformers, developed models to predict geomagnetic storms and solar flares before they happen (our work with ISRO won a best paper award), wrote classifiers for distant celestial bodies from sparse telescopic data in a collaboration with NASA JPL, and solved a (reasonably hard), unsolved partial differential equation that will (hopefully!) revolutionize how we calculate signed distance functions. My affiliations have fortunately been with some great and wonderful institutions, and this is an inexhaustive list of places I’ve worked with actively for at least 2 months: Caltech (2023), NASA JPL (2023-24), MIT (2024), ISRO (2023), DRDO (2022), IIT Gandhinagar (2020-24).

To start off, here are some blogs that detail my work. While most are hand-written, some of these are assisted with GPT, summarizing whitepapers I’ve already written. diloco - Ergodic research

• to pull apart a differential equation: A piece on how differential equations arise in the analysis of neural networks, how differential equations represented via neural networks can be a powerful and principled way to think of solving a wide array of problems, and on my time working with the cutting-edge of neural operator learning. In some sense, this is also my undergraduate thesis.
• speedrunning a small Mixture of Experts: I was given private access to a 32\(\times\) MI250X cluster during my time at Microsoft, for a period of 6 months, during which time I learned to pretrain and posttrain my own SLMs, speedran NanoGPT a few times (but never made the world record), observed the scaling laws for myself, and tried to train my way up from G-Shard Transformers/Switch Transformers to proper aux-loss-free MOEs, wrote Flash Attention kernels for ROCM, and more. This blog serves as a distillation of my learnings in that time.
• the demotheosis manifesto: Opinions on civilization, humanity, and our future, from a day I felt very hopeful about the Great Human Enterprise.
• tales from the borderline: Essays from growing up in Agartala, Tripura, a landlocked little refugee town at the literal border of India.

My Background, in Gratuitous Detail

I am a half-baked computer scientist² and a sloppy, aspiring adventurer³. By most definitions, I would qualify as a (novice) machine learning researcher; I have worked for, published with, or consulted for, the institutions that follow, grouped roughly by the expectations set up during my engagement. To know in more detail about the nature of my involvement, you can hover over the links, or even click on them for further detail; feel free to peel the layers of the onion, as they say⁴.

Education

• iit gandhinagar (2020-24), wherein I did my undergraduate in Computer Science & Engineering; I graduated at the top of my class, and was thus awarded the President’s Gold Medal. My education was supported by the Cargill Global Scholarship.

Academic Research

• Caltech & Jet Propulsion Laboratory, NASA (Summer, 2023) wherein I interned as a Caltech SURF student under Prof. Ashish Mahabal and Dr. Nitin Singh towards problems in (a) unsupervised classification of celestial bodies from highly noisy, sparse lightcurves; (b) built the world’s largest (as of 2023) database of bacterial genomic and phenotypic data, towards problems in spacecraft sanitization.
• Indian Space Research Organization (ISRO) (Fall, 2023) wherein our work on predicting geomagnetic storms before they struck the earth won the Best Paper Award at IEEE SPACE 2025, among submissions from all major space/defense research organizations in the subcontinent.
• Defense Research and Development Organization (DRDO) (Summer, 2022) wherein I worked with a team on various preliminary problems towards coordination in drone swarms⁵.
• Caltech & Tensorlab (Fall, 2025) wherein I solved a standing problem in deriving a closed-form Green’s function for the screened poisson equation in free-space for \(\mathbb{R}^n\), generalizing previously known solutions in 2D & 3D to arbitrary dimensions.

Industrial Research

• Mudrex Capital (Winter, 2023) wherein I worked as a Quantitative Developer, which is another way of saying I was an Algorithmic Trading Engineer. I developed a backtesting library there (part of which I opensourced) and saw the market beat my belief in DNNs as the answer to everything into submission.
• Microsoft AI (2024-2025) wherein I as an Applied Scientist and worked primarily on training cross-encoders on very large datasets, and got to see the scaling laws in front of my eyes. It was also when I learned to pretrain and stabilized my first large Mixture of Experts transformer, which was a long-term goal.

This is also a place where I wish to lay down a taxonomy of ideas that I have recently started to group under a single header: Demotheosis, a neologism combining the Greek demo, from demos, meaning “people”, and theosis from theos (relating to God), which by way of apotheosis means here to “become divine”. In short, Demotheosis is a masthead for ideas around how we, as a civilization, begin our conquest of the stars, and as a result, become a civilization that is no less than the Olympians we invented millenia ago.

Spain was a two-month voyage away from the Americas when Christopher Columbus made his first voyage, and began the cycle of colonialism. Today, the Parker Solar Probe reaches the perihelion of the Sun in three months! With an explosion of compute, and a subsequent need for energy acting as a forcing function for expedition (much like spice in the time of Columbus), it is hard to imagine that the seeds of a proto-Dyson Swarm is too far away. We are committed, we intrinsically have skin in the game, and the idea of a mesh of solar sails siphoning the Sun’s energy into training intelligences that far surpass our own recedes fast from the annals of science fiction, into something that one day will be scientific history. I posit that we must see machine intelligence as a way of storing compute – and hence, somewhat abstractly, as a way of storing the promise of harnessable energy.

Recent & Notable

• On Software Complexity — Why modern software is so complicated
• The Future of Programming — Speculations on how programming will evolve
• Notes on Scientific Method — How science actually works
• On Long-term Thinking — Why we should think in decades and centuries

Sections

• Essays — Longer pieces exploring ideas in depth
• Books — Reviews and notes on books I’ve read
• Links — Interesting things from around the web
• About — About this site and me

What is This?

This site is inspired by Gwern.net. Like Gwern’s site, I aim to:

• Write content that improves over time
• Use confidence tags and importance ratings
• Build an interconnected web of ideas
• Embrace open access and transparency

Pages include status markers showing their maturity. Some are seedlings; others are well-developed. The garden grows through constant tending. ## Field Notes

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Integer placerat, odio eu gravida pretium, nunc mi scelerisque magna, in interdum mauris ipsum ac erat. Quisque dictum urna nisi, id sagittis ex ullamcorper a. Sed eu sapien id eros aliquet rhoncus. Proin sit amet leo interdum, fringilla neque nec, porttitor erat. Phasellus et aliquam diam. Praesent laoreet augue eget aliquet dignissim. Fusce imperdiet porttitor molestie. Donec a est ligula. Integer rutrum, purus ut dictum laoreet, nisl leo vehicula arcu, ac vehicula diam lacus sit amet nunc. Lorem ipsum dolor sit amet consectetur adipisicing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.

Curabitur vulputate lectus non sapien consectetur, nec interdum nunc sollicitudin. Pellentesque habitant morbi tristique senectus et netus et malesuada fames ac turpis egestas. Vivamus faucibus nibh vitae quam facilisis auctor. Suspendisse nec semper arcu. Aenean vestibulum ligula ligula, eu fermentum metus pharetra sit amet. Duis mauris lorem, venenatis gravida nisi eget, ullamcorper dictum nibh. Nulla facilisi. Pellentesque dignissim consectetur sem, sed viverra nunc pretium quis. Integer rutrum tortor quis finibus iaculis. Quisque lacus augue, semper vel justo nec, ornare laoreet nulla. Pellentesque nibh mi, luctus ut eros sed, imperdiet condimentum orci.

“Retro computing aesthetics are a reminder that typography can feel tactile even when it’s purely digital.”

Mauris tincidunt tellus a felis pharetra, vel fermentum lacus aliquam. Vestibulum nec mi eget dui lobortis feugiat vel vitae magna. Quisque quis dui ex. Mauris accumsan turpis id tortor porttitor lobortis. Donec viverra diam ac porttitor faucibus. Maecenas eget metus id sem convallis luctus. Fusce sollicitudin nisl id orci accumsan, ut facilisis erat mattis. Aliquam facilisis eu lectus a pretium. Cras ut convallis arcu. Nam sodales aliquam velit, eget facilisis lacus volutpat quis.

Research Streams

The garden spans multiple lines of inquiry. Each column highlights a track I come back to repeatedly while drafting essays or notes.

Contact

Feel free to reach out—I appreciate corrections, suggestions, and interesting discussions. See the about page for details.

---

1. Primarily, I am trying to look at tensor parallelism for training foundational models over a swarm of edge devices. Solar-powered AI satellite must communicate with each other, after all!↩︎

2. In that I haven’t done a PhD yet, which people seem to point out is a requisite for calling myself a scientist. I disagree, but I still wouldn’t call myself a computer scientist, out of respect for the subject. I am to the computer scientist what an electrician is to the electrical engineer.↩︎

3. I come from a family of folk theatre artistes, brought up in the borders of India. I have travelled the length and breadth of the country, as part of the troupe as a child, and solo as a new adult. One picks up a few good stories along the way.↩︎

4. I am (as of this writing) freshly 24: that age which, a century ago, drove men to war, today drives me to claim this piece of real estate on the internet, fashion it shamelessly after one of my favourite internet haunts, and decide to start writing about the things I find at least mildly interesting, sometimes deeply fascinating, and almost always not in need of my opinion.↩︎

5. A vast majority of this is under NDA, but we did work on the problem over 6 months and had some interesting results. ↩︎