I had no precognition I’d be doing ML research today.

Hi! I’m Varshita. I graduated from IIIT Hyderabad this year (2025) as a CND (B.Tech in Computer Science and Master of Science in Computational Natural Sciences by Research) student after a very meaningful but less-than-a-year stint as a dual degree student at Precog. My time at Precog, however, goes way back to Monsoon 2022.

Precogger by proximity

My introduction to the Precog lab’s culture began by accident, co-working alongside some of its members and my friends in the recently vacated and under-renovation T-Hub space -It was a trove of playful interiors, couches that rats loved to snack on, glass whiteboards, and panoramic windows that let in the thick monsoon winds (with a faithful accompaniment of mosquitos of course). This was my absolute favourite place on campus. I’d be remiss if I didn’t mention the surreal sunsets and sunrises. 

At this time, I wasn’t in Precog. In fact, I wasn’t working on machine learning/deep learning at all, and I’d only taken a basic ML course. But, this period was my ‘real’ introduction to ML research and its day-to-days: Conversations, and mostly existing and doing my own work around loud discussions, (crappy) jokes about (crappy) results, and random whiteboarding sessions. 

I was dabbling in neuroscience at that time, and I saw an awful lot of parallels with ML: ‘Wait, this is neuroscience but you can just cut open the model and do whatever with it?!’’. I loved the extreme hackability of the models without the high barrier of entry that came with physical experiments. I was also surprised at how accessible and ‘close to surface’ even frontier ML research can be (Ilya Sutskever perfectly articulates this quite well here): I could understand a lot of the discussions without any formal introduction to the domain.

Beyond academics, this was also culturally novel to me. I’d never been in a room where research was discussed with such excitement – I’d mostly seen it treated as a chore, something done with little ownership and even less freedom to think for oneself. It was positively strange to see my friends not needing to report every minor design choice to PK; he simply trusted their capability to think independently. In fact, Precog’s research topics are largely dependent on its students, and I’m excited to see how it shapes up in the future!

TAing the Responsible and Safe AI Course

I had been engaging with AI safety literature on my own, so when PK announced a course on the topic, I knew I could contribute meaningfully as a TA. I had good familiarity with conceptual and philosophical aspects of current AI safety discourse- the “taste-building” parts of the curriculum. My excitement, however, was quickly followed by a healthy dose of a very specific terror: As a TA, our job went way beyond high-level discussions. We had to design the assignments, mentor projects, and conduct live viva evaluations where we’d probe students’ understanding of their own code. 

What if a student implemented an attack like GCG, and I didn’t grasp the implementation details myself? The thought of being unable to judge their work, or worse, having to offload the technical parts of my responsibilities to the other TAs, was as mortifying as it was motivating. As it turned out, this would finally force me to get my hands dirty and fight with the specifics – the code, the math, and whatever else needed doing. In hindsight, my background also uniquely positioned me to provide hands-on help, as I was so recently and intimately acquainted with many of the same doubts that the students had.

Tale as old as time, but personally verified: teaching someone else is the most unforgiving test of whether you actually understand something or just think you do.

A bit late to the party, but officially in Precog!

After a semester of TAing the course, at the end of my 4th year and despite the obvious risks (such as delayed graduation) of switching advisors late in my degree, I applied to work with PK as a dual degree student. I believed I could make significant progress here – a product of my interest, the lab culture I’d observed, the support infrastructure, and the accessibility of the field.

I will be forever grateful that he gave me the opportunity despite the unusual timeline!

The transition wasn’t always easy, but I felt very supported. The lab folk (ones I was already friends with, and ones I became friends with eventually) were incredibly kind and helped me ramp up quickly, and also made this period one of the best times of my college! I was also deeply grateful for PK, who made it a point to check on my personal well-being, not just my project’s progress – that meant a lot 🙂

Beyond our project meetings, we had WU’s (What’s Up). 3 hour long WUs. For a while, the length of these meetings didn’t make sense to me. But in hindsight these long-form discussions with people not on project were invaluable for killing bad ideas early, avoiding fatal mistakes, and sometimes resulted in major pivots. As I was new to ML (especially research), I learned enormously from these sessions – from direct feedback on my work, and from the thought process that went into each project, watching how fellow labmates provided critique. The questions people asked taught me how to question my own work. Some of our best research developments came from questioning assumptions made in the existing literature, and first led to our early workshop paper ‘Sanity Checks for Graph Unlearning’.

Doing the ‘research’ was one thing, but peer review was its own education.

Our work was first rejected from ICLR, and frankly, I was deflated and a bit pessimistic about submitting to ICML. But my teammates and PK had more conviction. They pushed to aim high again, and PK assured me that my graduation wouldn’t be held hostage by a paper rejection. For ICML, we made the specific changes that we felt would best build trust in our central thesis. We got lukewarm, borderline scores, so we poured days of work into the rebuttal. And then… surprise, surprise: Silence. Our reviewers didn’t engage. I was again reminded that a paper’s fate also hinges on aspects that aren’t related to its scientific merit.

While we were awaiting these results, I had the chance to attend ICLR in Singapore. It was a strangely clarifying experience. As I walked around the poster session, I thought, “It would make sense for our paper to be at a venue like this”. With improvements to the ICML paper, I just hoped the ‘variance’ was in our favour this time. A few weeks later, our paper was accepted. Personally, I’d count that self-realisation to be nearly as big of a win, because such ‘on-paper’ wins are sparse even if you’ve done good work. It’s worth mentioning that these not-on-paper wins were celebrated in Precog too. Our arXiv paper, and the creation of our website was met with as much excitement as when we received the acceptance. 

Next Up: New Beginn(er)ings

I’ve started as a pre-doctoral researcher at Google DeepMind.

Guess it’s time to start obsessing over research until 4 a.m. again (or whatever the healthy-working-adult equivalent of that is).

For preparing me for that, and for everything else, again, thank you, Precog!