We were cruising at a comfortable 50 miles per hour on San Diego’s I5 as we came up on a bend. I intuitively held on to the seat in front of me, as I was filming from the back seat. George was sitting comfortably, with his hands on his lap, talking to the rest of us about what’s in store for Comma.ai. The car followed the lane without breaking a sweat. “I can’t believe this is possible. It’s a phone on the windshield!”, I said. “Yup”, replied George. “And there are about 10 of us in the company”.
I spent this summer in sunny San Diego, interning at Comma.ai as part of the research team. This essay is about my major takeaways from the incredible 3 months I spent here. I first heard of Comma from an interview George Hotz (founder of Comma) did with Lex Fridman. In February, I found out on twitter that Comma was hosting a hackathon. Being on a good hackathon winning streak then, I applied and ended up at their office. It was exciting and exhausting - as hackathons typically are - and at the end of it, I had an internship offer. This was by far the most fun and ‘out there’ internship ‘interview’ process I had been through, and accepting the offer was a no-brainer.
Comma is unlike any company I’ve ever seen. “A company like Comma shouldn’t exist, yet here we are”, was a recent tweet from the Comma handle. This can potentially be construed as a carefully crafted and sanitized PR facade, given how most companies want a bubbly and woke twitter presence. But it is hilarious how simply true the tweet is. For the uninitiated, Comma is a company of ~10 people building openpilot - the advanced driver assistant system that runs on a phone-like hardware, sits on the windshield and takes over your car when you want it to (most popular car models are supported). All openpilot code is open-sourced, on GitHub. All ML models are freely available on GitHub. Comma is already profitable. If you’re even remotely familiar with the autonomous driving industry, you must be in the middle of a laughing fit.
A company like Comma shouldn’t exist, yet here we are.
We at the research team are responsible for producing models that ‘drive’. The problems we’re trying to solve are very hard and mostly are not problems that have clear answers. More often than not, they involve figuring out clever ways to process millions of miles of driving data we have, and generate ground truths to eventually train models on. One of the big projects I worked on was to estimate calibration parameters for devices incorrectly mounted on windshields. This has a non trivial implication on driving behaviour. I spent a long time understanding the physics behind the phenomenon I was trying to model. Getting started, to be very frank, was as intimidating as it was exciting. To have access to a state of the art data-center to run my experiments felt strange, to the poor grad student in me. After a bunch of trials and errors, and discussions with teammates, I had a final method. It involves a least-squares plane fitting algorithm using the vision stream, the GPS stream, and the semantic segmentation engine. A model was thus trained. I was extremely thrilled at the sub-centimeter, sub-degree accuracy. But this was only half the task. The rest of problem was to make sure this got into production. My only regret is not iterating fast enough on some of the sub-projects I was working on. ‘End-to-end’ is a recurring theme at Comma - from models to people. Engineers own problems, and as much as possible, see them through to their ends. You generate data, you create your own tools, you build models, you put it in production, an you test it - even on a car. While this might seem like a bit much, I think it has cultivated a sense of skin-in-the-game for everyone. All employees drive with openpilot almost everyday. It’s a product engineers are building to delight themselves, everyday. And there’s something very special about that.
Now before you accuse me of being high on the Comma koolaid, I am the first to admit that Comma has its own issues. Which company doesn’t - let alone one with 10 people and a product used by thousands? Hiring is always hard - especially in domains where there are no ‘solved problems’. Comma is relatively unknown even amongst friends who I think would love such a place. This results in a very low bus factor - the number of people that can be thrown under a bus or off a cliff before a project stops - which is not ideal for any company. You always want enough leeway to throw a few people off the cliff, here and there :) Internal tooling, in the presence of engineers dedicated to its cause, would be more mature and not break often. Ideally, the best engineers would have all the drudgery taken off their plate, resulting in more time to wonder about the future of Comma.
To work at here, I think one needs unsubscribe from the tribes most of us naturally fall into. There are no lofty titles, there are no managers and there are certainly no skills to endorse on LinkedIn. There are no ladders to climb, there is no politicking to do - the structure of the company renders these things irrelevant. The people are extremely competent and extremely kind - a potent combination for a laid-back-yet-very-high-output environment. To truly enjoy working at Comma, one needs to have a deep sense of fascination for the pursuit of truth. I suspect it is very hard to work at Comma if you want ‘just another job’. The most thrilling and intimidating property of working at Comma is the freedom - because freedom is a double edged sword. But you will only enjoy working here if you would not have it any other way. There’s only one meeting each week and everyone’s priorities are for everyone else to see on a wiki. Comma does not remotely fit into the outline of a company. It is an experiment in freedom and in playful defiance - one that is felt the moment you enter the office to a “F/❤️ U ELON” license plate, it is felt in the free high quality software on Github, it is felt in conversations over (Masterchef quality, really!) lunch, or over a team surfing trip on a Friday afternoon - because why not?
When I left India exactly a year ago, I was determined to get into a bigger pond. I remember making a note on the flight that said something to the effect of ‘strive for competence, take risks, have fun’. I think this summer was a great step in that direction. Comma’s original motto was to “Make Driving Chill”. In one of the earliest meetings I attended, it was officially changed to “Solve Self Driving while Delivering Shippable Intermediaries”. This is the single biggest difference between Comma and most other companies trying to solve self driving. Misgivings of this industry will warrant an entire post in itself. But it suffices to say that people drive thousands of miles on openpilot everyday. Comma will take more and more and more actions away from those drivers, delightfully. Comma will eventually turn any car into a level 3 car - it is the android, to Tesla’s iOS. But most importantly, Comma will make driving chill.
Do checkout Comma’s career page if what you read above feels like your cup of tea!