← Atharva Naik

Blog

Posts

November 17, 2026 · 3 min read

What I Would Work On If I Were Starting My PhD Today

My reply to an incoming PhD student, assuming some form of AGI arrives within a PhD: applications VCs won't fund yet, auditing and understanding AI-driven research, and making sure the benefits reach speakers of every language.

PhD advice, research directions, AGI
November 14, 2026 · 7 min read

In Defense of the "Research Blog"

Research blogs get dismissed as less rigorous than papers, but arXiv preprints guarantee no more rigor or reproducibility. Blogs are a complementary, smaller unit of research communication with their own uses.

research process, peer review
Auto Weight-Space Programming · Part 1November 13, 2026 · 30 min read

Defining the Problem and Initial Experiments

with Emmy Liu and Varun Gangal

We formalize weight-space programming and set the goal of auto-WSP: a model reading and editing its own weights. As a toy first step, Qwen3.5-4B learns to apply the logit lens to 128-number slices of its own unseen neurons.

interpretability, weight-space programming, SFT
Defining the Problem and Initial Experiments

Series

Auto Weight-Space Programming

Training open models to read and write their own weights.

Elsewhere

co-authored[external]

Emergent (Mis)Alignment: Exploring the Hidden Link Between Code Security and AI Alignment

with Abhinav Rao, Alex Xie, Anmol Agarwal, Shubham Gandhi, Michael Hilton and Carolyn Rosé (CMU)