Projects

nanogpt world record (in review)

Achieved a new world record on the NanoGPT speedrun leaderboard using test-time training, reducing training time from 99.3s to 95.9s to 3.28 cross-entropy loss on the FineWeb validation set.

induction and inquiry via probabilistic reasoning over language and code

Human inductive reasoning is best explained by sequential probabilistic reasoning over mental programs — a mix of natural language and source code. Our LLM-augmented Sequential Monte Carlo model captures garden-pathing, anchoring, and human-like active inquiry.

reasoning research @ stealth lab

master's thesis

People are incredibly flexible and efficient inductive reasoners. How can we model this ability computationally? In my thesis, I propose two hypotheses: people may operate over an incredibly vast language which is made tractable via a strong bottom-up proposal model, or people may learn task-specific languages to reason over.

resource-rational task decomposition with theory of mind

Compositionality, the capacity to understand and generate complex ideas by combining simpler ones, is a critical attribute of human intelligence and plays a pivotal role in sophisticated problem-solving tasks.

hyperbolic vq-vaes

Combines vector-quantization with hyperbolic embedding space to train a Hyperbolic Vector-quantized Variational Autoencoder towards learning discrete hierarchical representation.

how to establish if young infants comprehend compositionality

Prior work has shown that 4-year-olds can use compositionality in visual tasks, but failed to show the same ability in younger infants. Designed a study to test this ability.

gabor-constrained neural networks for transfer learning

We explore whether constraining a machine vision system to be more human-like improves the performance of the system in learning new tasks.

larc

LARC (Language-annotated Abstraction and Reasoning Corpus) is a dataset I collected with Yewen Pu and other post-doctorates in the MIT Computational Cognitive Sciences lab.

generative art: one a day

Art created using code. Varies from combining GANs and diffusion models with CLIP to generating simple flow fields.

role of individual neurons in multi-modal models

Explored how individual neurons contribute to the final answer of large multimodal models, and how model explainability relates to performance.

inferring moral values via hierarchical bayesian modeling

Explored computationally inferring the moral values of different cultures using a hierarchical Bayesian model on The Moral Machine domain.

acquacchi

A quarantine project to learn C. A strong chess engine written in C, ported to Javascript. It expands alpha-beta search with heuristics and bit-level manipulations.

kandula

API and app that enables anyone to create and customize stock-trading algorithms without code, just by tuning parameters.

lithium-oxygen battery redox mediators

We looked at using different solvent mixtures to tune the stability of inorganic redox mediators in Li-O2 batteries.