There is a giant elephant in the room that grew up from a young calf all the way to a giant mammoth in the span of about a year - LLM based generative AI models. It’s unreal the amount of progress they have made over the past few months. They went from barely being able to solve codeforces division 1 level problems to literally being used in the frontier of theoretical research – across math, physics and especially my field : quantum information. The recent crack at Navier Stokes was just a cherry on top.

It has lead to a lot of mathematicians getting their feelings hurt : link, another link. But this knee jerk reaction to AI mostly feels like a coping mechanism. I have seen countless posts actually criticizing the help AI provides in doing research, the above links included, out of fear of being replaced. There are literally university professors thinking of not sharing their brilliant ideas or research directions with the community out of fear of getting scooped by AI. If anything, this sheds light on the way academia is (was?) structured : these people aren’t excited about knowing more or doing more. They are sad and afraid that they won’t be credited for generating that knowledge. Really tells you exactly whose intentions are truly misaligned.

Not to bash on them too hard : some of their concerns about training young researchers are well founded. And in their defence, AI companies are no angels (please don’t cancel my subscription.) But there is an implicit assumption that the old way of doing things was the best and the only way. If academics want to restrict themselves and their subordinates to be “AI-free”, all the power to them. But that in itself shouldn’t give them a sense of superiority over other AI-consulting mortals. You have to be a fool not to foresee the potential for accelerating research using this technology.

I’m using AI very liberally for my research. It’s an amazing sparring partner to bounce ideas off of. Moreover, it’s a super-human search engine for ideas. My latest paper “Challenges in Barren Plateau Mitigation with Dynamic Parameterized Quantum Circuits” [arxiv] benefited a lot from Claude code programming a symbolic manipulation engine. It also helped me out with some basic matrix inequalities and variance decomposition formulas that otherwise would have taken me days to research on my own, not to mention paywalls. Claude Sonnet wasn’t as good as Fable/Mythos but it did a decent job.

And it got accepted as a contributed talk at QTML 2026! It’s also currently with the editors of Quantum Journal. The paper itself explores a line of recent works in barren plateau mitigation in dynamic circuits. Specifically, we show that constructions like the one in this paper have some serious issues. (Fun fact: I had emailed one of its authors for an internship but got ghosted. So when I found out these issues I was very happy to publish them.) As for analyzing dynamic circuit gadgets : I just saw CPTP channels and went

Recorded footage of me realising how to model dynamic circuit gadgets. (Don't message me with anime questions I have never seen this. I just found the meme online.)

This paper started out as an attempt to solve the BP problem. But in itself, this was a negative result. Nevertheless, with this acceptance as encouragement, I want to proceed to solve other challenging problems. I’m currently working on super fast and cheap exact matching decoders for surface codes and Claude code has been invaluable for me to run experiments, debug and optimize my algorithms. I’ll make sure to credit it exactly for its contributions. We’ll be submitting our work to ISCA and I am confident it will get accepted. It quite literally mogs last year’s ISCA decoder paper.

As of now, Claude is not as great at setting research agendas or even writing a decent paper by itself. And it does produce a lot of slop – I define slop as anything nobody cares about. Moreover, an automated loop that feeds on slop produces a slop-cycle which may lead it down a rabbit hole, which is why it needs constant guidance; similar to how graduate students need guidance from their supervisors. But I expect it will only improve from here. All in all, I’m excited for the future the clankers will usher in!