Going to NeurIPS 2025!

I’ll be in San Diego for NeurIPS this year. I considered going a few times in grad school, but this is my first time making it.

I’ve often thought about how my PhD fell between two landmarks in AI: DeepMind’s AlphaGo beat the world Go champion right as I was starting out in 2016, and I graduated in 2022, right before ChatGPT came out and helped kick off the ongoing AI hype cycle. It really is shocking how rapidly the field has advanced - I’ll admit that I would not have expected the current capabilities of frontier AI models, despite having what was essentially a front-row seat to work at MIT and elsewhere.

Rather than continue on the path of neural network research I had started during my PhD, I decided to pursue entrepreneurship instead. I do think this gave me a lot of practical skills in leadership, product development, and building a company, but I’ve been out of the game for over 3 years now. A lot has changed, or so I thought. In reality, most frontier AI models rely on the transformer architecture, which is from 2017! The stunning accomplishments of LLMs are due mostly the increased scale at which models are trained: more data, and more compute power (there’s a reason Nvidia’s stock price has done so well the past few years). From the limited amount of catching up I’ve done recently, it seems like there are still a lot of fundamental questions about how these models work.

I’m looking forward to learning more at NeurIPS this week, and I will try to compile more notes and thoughts into additional posts here.