Highlights:

See full resume here

EDUCATION

  • The University of Texas at Austin, Austin, TX
    • Degree: Bachelor of Computer Science
    • Program: Turing Scholars Honors Program
    • GPA: 3.9652
    • Graduation Date: May 2025

RESEARCH

  • Explorations of Self-Repair in Language Models [arxiv] [ tweet thread]
    • Cody Rushing, Neel Nanda
    • Accepted to ICML 2024; Accepted to SeT LLM @ ICLR 2024 Workshop | Oral
  • Copy Suppression: Comprehensively Understanding an Attention Head [arxiv] [blog] [streamlit]
    • Callum McDougall*, Arthur Conmy*, Cody Rushing*, Thomas McGrath, Neel Nanda
    • Accepted to NeurIPS ATTRIB 2023 Workshop

EXPERIENCE

  • ML Alignment Theory Scholars Program
    • Mechanistic Interpretability Researcher
      • Full-time: Berkeley, CA; May 2023-August 2023
      • Part-time: Austin, TX; October 2023 - January 2024
      • Key Achievements:
        • First Authored Copy Suppression Mechanistic Interpretability paper explaining 76.9% of an LLM Attention Head
        • Full-time research mentorship from Neel Nanda (Deepmind) on mechanistic interpretability
        • Reverse-engineered preliminary circuitry for the completion of ‘dual pairs’ of words in GPT-2 Small.
  • CEC Entertainment
    • Cybersecurity Intern
      • Full-time: Irving, Tx; July 2021 - August 2021
      • Responsibilities:
        • Operated Vulnerability Scanning and Penetration Testing software (Nessus, Wireshark, Metasploit, Burp Suite)
        • Systematized Employee Equipment Imaging and Established ~35 New Employee Stations

PROJECTS

  • “Mini Shakespeare”
    • Description: Built a decoder-only transformer from scratch with PyTorch, and then trained it on a partial Shakespeare corpus; combined with decoding methods, it generates ‘Shakespearean’ dialogue
  • Goal Conditioned RL Agent
      Description: Implemented a Goal Conditioned RL Agent to solve custom BabyAI environments report, code,
  • Multicycle Verilog Processor Pipeline
    • Description: Implemented multicycle Verilog processor pipeline, optimized with Perceptron-based branch prediction for highly efficient control flow (~33% mean cycle count improvement)
  • Custom Java File Encoder/Decoder
    • Description: Assembled custom Java file encoder/decoder, largely based on Huffman Encoding
  • Replicating Scaling Laws
    • Description: Authored report automating training of six variable Convolutional Neural Networks through wandb to extrapolate scaling laws
  • CS439H Project
    • Description: Class-wide collaboration to build Operating System Kernel which runs networked multiplayer Doom; includes events, virtual memory, user/kernel preemption, signals, file descriptors, etc.
  • “An Unhackable Internet”
    • Description: Directed educational video characterizing the Quantum Internet; 60+ hour, placed top 5% of videos in competition