David  D.  Baek

I am currently a second-year PhD student at MIT EECS, where I am fortunate to be advised by Prof. Max Tegmark in the Tegmark AI Safety Group. I am passionate about leveraging mathematical analysis, machine learning, and algorithm design to solve challenging real-world problems!

Email  /  CV  /  Google Scholar  /  LinkedIn  /  Twitter  

Research

My research interest revolves around representation learning, LLM interpretability, and AI alignment. Modern machine learning models often operate like a black-box: my goal is to understand the mechanism that enabled the success of these large-scale models, thereby making them more interpretable and aligned with human's goals.

Harmonic Loss Trains Interpretable AI Models

We introduce harmonic loss as alternative to the standard CE loss for training neural networks and LLMs. The post about the paper has received 1.1 million views on twitter.

The Geometry of Concepts: Sparse Autoencoder Feature Structure

We study geometrical structure of Sparse Autoencoder feature vectors at three scales: (a) “atomic” small-scale, (b) “brain” intermediate-scale, and (c) “galaxy” large-scale. The post about the paper has received 300K views on twitter.

Generalization from Starvation: Hints of Universality in LLM Knowledge Graph Learning

We study geometrical structure of knowledge representations in Large Language Models (LLMs), and examine hints of representation universality.

GenEFT: Understanding Statics and Dynamics of Model Generalization via Effective Theory

We present GenEFT: an effective theory framework for shedding light on the statics and dynamics of neural network generalization, and illustrate it with graph learning examples.

Design of scalable superconducting quantum circuits using flip-chip assembly

We present a method for designing highly scalable superconducting quantum circuits using flip-chip assembly and multi-path coupling strategy.

Class Projects

Filtering Data to Improve NeRF for Driving Scenes

We propose a filter-based data-preprocessing strategy for NeRF that enables autonomous driving scene reconstruction with reduced compute time and memory, with the minimal performance loss possible.

Automated System for effectively managing leaves of Korean soldiers

We built a web application that automatically manages the leaves of Korean soldiers, using Node.js, Express, MongoDB, Passport.js for backend and Vue, Vuetify and Chart.js for the frontend.

Wanderlust: Community for Hikers

We built a mobile application which serves as both a guide and a community to hikers all over the world. We used Flutter, Google Maps, and Google Firebase to build our application.

Design and source code from Leonid Keselman's Jekyll fork of Jon Barron's website.