Deciphering cool German texts using LLMs with Prof. Thomas Lacqueur

Playing with Lil Chungus (Mouse 103) and training ML models on time-series data

Neuromorphic Project: Spikeformers

Investigated efficiency mechanisms for Spiking Vision Transformers, implementing temporal early exit (SEENN-I) and spatial token pruning. Achieved 36% compute reduction with under 1% accuracy loss on CIFAR-10.

Working on MSN (via Neurotech @ Berkeley)

Conducting independent research on the limitations of Reinforcement Learning with Human Feedback (RLHF), including value pluralism, reward hacking, and the impossibility theorem of fairness in inherently biased large-scale human-sourced datasets.

Developing a theoretical framework for internalized value learning, where AI agents learn from human datasets while evaluating information through an internal system of ethics and objectives, enabling moral reasoning rather than mere imitation of observed behavior

Surveying existing literature in RL, cognitive science, and moral philosophy, and connecting them to alignment mechanisms. Collaborating with Google engineer on a blog-style project to communicate technical research insights to a broader audience.

Current Research:

1."Reinforcement Learning through Human Feedback (RLHF) may not be enough"

(Presented at CITRIS and Banatao Institute Tech Policy Symposium, paper coming soon)

2. Investigation of Emotion in RL