about

Hi, I’m Ellen.

I like solving puzzles and buildings things.

From 2023-2025, I worked on ml car stuff at comma.ai and Tesla FSD. At comma.ai I fell in love with the culture of hacking cars and making open-source software people actually use and love. At Tesla I had a lot of fun I shipping models to millions of users and getting to launch Robotaxi in Austin!

Talk to me about:

• sim2real diffusion Push-T on $200 Lerobot Koch arms in my dorm room
• building an autonomous navigation algorithm in <48 hrs which successfully navigated around the comma office e2e camera-only (shoutout to my seriously amazing teammates!)
• AlphaZero training for self-driving/robotics (and where AlphaZero does or does not work for self-driving)
• improving System 2 reasoning for multimodal models at Tesla (when going on “autopilot” isn’t sufficient)
• learning car controls through evolution
• exploration and RL through set theoretical frameworks
• CNN-based screening tool deployed/tested with doctors (why deployment and real-world feedback matters for ML!)¹

Some other cool things I’ve done: 10+ years of competitive saber fencing (this was a huge part of my life - at my peak won 5th place in Romania world cup for U17 age group!) and fenced for Stanford’s Div I team (2024-2025), spent a summer learning to fly (40+ hours on Cessna 172N, but PPL is currently on pause…), studied oil painting at an art institute in Florence and spoke poor Italian (seeing the world outside the silicon valley bubble :P), won a hackathon with friends and got to fly a R22 around SF (yes I am slowly collecting all the modes of transport…), getting into blue vs red teaming and cyber stuff in high school and being a part of the CyberPatriot XIII national championship team, running marathons (including an impromptu Stanford2Sea), making stuff (wooden katana, laser-cut neon sign, 3d printed my brain from MRI scan, geodesic dome…)

Some questions I’m currently thinking about: why can’t LLMs in their current form get parity with AlphaGo? Why are LLMs so notoriously bad at learning to play games? What is the right metric (similar to interventions/mi) to scale for robotics (interventions/time for task?) In what sense is an LLM a “world model”, and in what sense is it reasonable to think about LLMs taking actions? If humanoid hype is like self-driving where people tend to severely underestimate the timeline, how to avoid the trap of false progress and build something real and sustainable? When will we get child AI where we start with a base model that can learn new skills, be curious, and learn on its own from interaction? What’s the best way to augment human intelligence and climb up the skill ladder as it rises?

chat with me! (☞ﾟヮﾟ)☞ ellenjxu@stanford.edu

---

1. … as well as the best roofs around Stanford, my ranking of top Italian cheeses (beware I will judge you based on your cheese preferences), gardening tips for Arch Linux (compositors are overrated), favorite puzzles (test-solve ones I’ve made or collected from friends), the long list of books in my to-read and things I want to write² ↩

2. When I was growing up I wanted to be an author + inventor. I still think da Vinci is the coolest person ever. ↩