From 2016 - 2019, I was a machine learning engineer at Uber ATG (Advanced Technologies Group), Uber’s R&D division for self-driving cars. I worked on 2 teams : deep-learning and motion planning.
🧑💻 My technical work at Uber
Although I cannot publicly share the work I did at Uber, I can share some of the ideas I worked on. Here are notes in ML and optimization that I created which were related to my work at Uber.
🤖 Robotics & optimization
- Linear Quadratic Regulator - optimal control linear dynamics and quadratic costs
- DDP / iLQR - optimal control for nonlinear dynamics and quadratic costs
🧮 Machine Learning / AI
- Variational Inference - a method for approximating intractable posterior distributions
- Gradient Descent Methods - methods for picking the descent direction
- Taylor Series Approximation - methods for approximating functions
- Reinforcement Learning - a framework for optimizing an unknown reward function
💻 Software Engineering
- ML Infrastructure Design - a design for a machine learning infrastructure
🤔 Takeaways
Here are some quick reflections from my time at Uber ATG.
1️⃣ Premature scaling at Uber ATG might’ve killed us
At Uber ATG, one of the leadership missteps was trying to scale our self-driving program too quickly. We had hundreds of cars on the streets of Pittsburgh, Phoenix, and San Francisco, burning millions of dollars daily. Leadership justified this by saying we needed the data and had to prepare for scaling. But the truth was, we didn’t yet have a product that could safely handle urban driving. Premature scaling shifted our focus to maintenance over innovation and drained resources. In hindsight, we should have kept far fewer cars on the road and invested more in R&D. While I didn’t have full visibility as I wasn’t in leadership, I believe this was a significant error.
2️⃣ How to pitch an opportunity that doesn’t exist
Uber ATG was my dream job in college because the job would involve advanced math and building one of the most impactful technologies of our time. Yet, when I was applying in the summer of 2016, I noticed that most of the people that worked at Uber ATG had robotics PhDs from places like CMU. There was no way I could compete, especially as an undergraduate with no publications or impressive work history.
One of the best things I did was to NOT ask for a full-time job and instead ask for a 1-year internship. I knew this was a better win-win because it lowered the interview criteria from “can this candidate contribute on day 1” to “can this candidate learn quickly and contribute meaningfully after a few months”? This internship technically didn’t exist on paper, but I pitched it to the head of ML at Uber ATG and this ultimately led to an offer.
Although not an Uber-specific lesson, this experience taught me that be creative in your asks and look for opportunities of win-win that might not be explicitly stated. Here’s the email I sent to Jeff that ultimatley led to my first job out of college!
