Since the summer of 2025, I have been working as a Machine Learning Research Intern at Cornell Tech. I’m grateful to be a part of Dr. Nikhil Garg’s AI Lab as the only undergraduate researcher, working with two PhD student researchers.
Though the work is continuously evolving and exploratory, the overall concept is to train sparse autoencoders (SAEs) on Bluesky posts to generate interpretable, concept-level features to inform new applications and social science research down the line.
One direction my team explored was using the interpretable features to generate paths of information flow on over 2 million Bluesky posts. The idea is to allow users to browse and navigate through Bluesky intuitively and expansively.
Meanwhile, I was leading my own project creating a continual-learning framework for a new SAE. The goal is to train the SAE on new data incrementally without retraining from scratch. This is important because Bluesky’s content is constantly evolving, and we need to keep our model up-to-date with minimal computational cost. With this new SAE, we want to be able to capture emerging trends and patterns in social media.
Combining research and development and using PyTorch and Jupyter Notebook, I built a prototype trained on academic ArXiv papers over time that gave promising results.
I tracked the evolution of neuron 747, the neuron for blockchain, over time and observed its concept drift. Its sudden spike in mentions around 2015 lines up with the development of the rise in blockchain technology, and its slight decline after 2020 aligns with a decrease in popularity of blockchain in academic research after the pandemic. Tracking the topic evolution of the neuron, it shifted from very low-level, foundational technology concepts to more high-level applications. Finally, through freezing old neurons, I was able to achieve a signficantly lower concept drift in all neurons, which is crucial for maintaining the interpretability of our SAE over time.
This project is still very much ongoing, and any updates on progress will be posted here.
