from Guide to the Job Hunt on Mar 19, 2023

Why (and why not) work at Tesla

I did a short 4-month internship with the vision team at Tesla AutoPilot, under Andrej Karpathy. Despite being a part of an insanely productive and impactful team, I didn't return for full-time. Here are reasons why I regret leaving Tesla AutoPilot — as well as reasons why I don't.

Tesla is well-known for grueling work hours, and it's for good reason: Through a combination of high expectations and company culture, long hours are the norm. However, it's not all doom and gloom. These long hours make the team as productive and impactful as it is.

This post is written for several possible categories of readers — effectively, anyone with an avid interest in working culture at Tesla, in particular for the Autopilot team:

• You're considering a job at Autopilot. This is the primary audience, as I'll focus on pros and cons that affect your everyday workday. The post is organized around pros and cons, influenced by culture, people, and the working environment.
• You're curious what working at Autopilot is like. I'll provide some insight, but the focus is not a "day in the life of". Rather, the focus is on higher-order observations about your productivity and career progression.

Just like with Why (and why not) work at Meta, the focus is on everyday concerns, but I'll briefly mention two common objections to working at Tesla Autopilot:

• "I don't like Elon Musk." Working at Tesla doesn't mean you have to be or become an ardent Elon fan. In fact, there are many longstanding employees that share a healthy skepticism of Elon. As a leader, he certainly has both a fair share of flaws and insights. We'll dive into the implications later.
• "Tesla's Full Self-Driving is dangerous." FSD has both impressive technical advancements and clear room for improvement. However, just as with Meta, a controversial product could be more reason to join: Given there are millions of Teslas on the road, a vast majority with some form of autopilot engaged, joining and making the system safer is more impactful than critiquing from the outside — unless you're a regulator. Unlike Meta, your salary at least doesn't stem from advertising revenue.

With that said, these objections aren't the focus of this post. Instead, our focus is on practical on-the-job concerns for working at Tesla, beyond the principle of joining.

As I discussed in How to make big decisions., discussions for major decisions should be broken up into a) criteria and b) rankings. For each of the major points below, I'll a) lead with the criteria and b) present information that will help you rank Tesla for that criteria.

The work is the main pro.

The main pros for Tesla generally relate to the actual work you're doing — both the impact and your own productivity.

Pro #1. Highly-focused work environment. The office is plain, team meetings are sparse, and company activities — be it all-hands or happy hours — do not exist.

• Impact: This results in a near-perfect working environment. Said plainly, the work is all there is to focus on. This furthermore means you become the most productive you you've ever seen, as (1) your working hours are used more productively and (2) expanded working hours mean more contiguous chunks of work time. As one scientist had said, you can grow 2x more productive with just 1-2 extra hours a day. Taken altogether, this boosts your output significantly — more time, used more productively.
• Comparison: Meta was much more distracting, by comparison. The multitude of employee perks were certainly enjoyable, but a number of factors contributed to fragmented days: Company activities such as yoga classes were common; the office was designed effectively as a playground; and the company information portal was an internal version of Facebook that bombarded you with random corporate news.

Pro #2. Highly-focused team. The team had its eyes on the prize at all times — the next model drop. This meant that all projects, and even experimental runs, were all candidates for the next dropped model.

• Short-term-oriented: All work was focused on guaranteed wins within a short timespan — usually within a day and no more than a few days. Very few risks were taken blindly, and every risk was a calculated one. This did mean that "maybe" ideas were quickly shot down, in favor of nearly-guaranteed ideas.
• Impact: This effectively meant that longer-term bets, such as technical debt cleanup, were harder to justify. Research for papers would be impossible in this environment, even if it isn't explicitly disallowed. Timeframes were also drastically changed: We planned 2-3 days in advance. A week was "long-term". Anything longer than a week was risky, and we never planned more than a few weeks in advance.
• Comparison: This focus is a rarity. At Tesla, the team works closely together on the same model. However, at Meta for example, each person owns a different project, individually. There are certainly many participants in update meetings, but for the most part, just 1-2 people are actually doing the work. This is to ensure that each employee can be assessed for performance reviews, but this also means the team as a whole is not very focused.

Pro #3. Fast-paced team. The Autopilot team moved quickly and efficiently, making a nimble and highly-productive team. This was partly driven by senior scientists who had dual strengths in management and technical skills.

• Quick iteration: If I posted good news from overnight experiments in the morning, 2-3 others would have picked up my patch and launched more experiments by noon. Within a few days, models would accumulate several micro-improvements, and within a week, almost every scientist would have positive results.
• Centralized: The entire team sat within close proximity. The 200-person Autopilot team sat in one area of one floor. The 15 vision scientists furthermore sat within just 3 rows. There were no sister teams working in a different timezone; all the experts you needed to talk to sat within 50 paces. Rather than email, schedule a call, then chat next week, just walk on over for a quick chat.
• Impact: The team iterated especially quickly on what would've otherwise been a black-box and difficult process of model development. As a team member, this made contributing and collaborating especially easy: Report improvements and push commits for others. In turn, keep an eye out for positive experiments and cherry-pick promising ones. This meant day-to-day deadlines and frequent cooler conversations with other scientists.
• Comparison: Pretty much any other company works at a snail's pace compared to this — because of the systems put in place. We'll discuss this further below, but subpar tooling, red tape, and distributed redundant teams are the prime suspects. For example, as we discussed in Why (and why not) work at Meta, there are highly-similar teams such as "On-Device AI" and "AI On-Device".

Pro #4. World-class tools. Tooling was very mature, with centralized experiment management, customized visualization tools, and a mature data pipeline.

• Impact: The tooling was a massive boost to productivity. Launching jobs, debugging models locally, and visualizing were all easy to do. Every task in the critical pipeline was well-optimized and anything that wasn't well-optimized yet was simplified with adhoc bash scripts we'd pass around the team.
• Comparison: At Meta, tools such as our build system buck — a symptom of the monorepo — was slow to run and a big drain on productivity. Every python run.py took ~10 minutes to run if you were lucky. Other tools such as our GPU cluster was constantly error-prone. Taken altogether, there was tooling, but it was very unwieldy to use.

Pro #5. Real impact on lives, daily. The team constantly reviews incoming data for mistakes and successes alike. For the vision team, missing or false detections are immediately triaged, cleaned, and pushed into our validation or training sets.

• For example, the team regularly reviews moments that Automatic Emergency Braking (AEB) engaged. In short, you get daily reminders of the lives you saved. There are vehicles about to drive into people, off roads, and into rivers, had it not been for AEB. This is about as real and immediate of an impact you can have.
• Impact: The team takes it work seriously, and thanks to Tesla's working environment, there's almost always a sense of urgency in understanding, improving, and validating onboard models. If nothing else, this realization can keep you on your toes and always on. When it comes to diagnosing real-world model performance, the team is very serious.
• Comparison: Other applications of computer vision are more prolific, maybe cooler-sounding, but less impactful — be it camera filters or generating pretty pictures. Among the most serious, widely-deployed applications of vision is FaceID, which certainly improves convenience but is a bit far from lifesaving.

On a side note, it was additionally a pro that Tesla Autopilot was transitioning to a vision-only and radar-less system during my stint there. This was reflective of a general principle at Tesla, that vision is enough to achieve full self-driving. Whether or not this is true, it made sense for me — as a computer vision researcher — to go to a workplace that prioritized computer vision. This meant my work had the most impact at Tesla; at another company such as Waymo, which uses LiDAR, my work would not have as great an impact, for a LiDAR-centric sensory pipeline.

Cons

Con #1. There is no "life" in "work-life balance". This is taken to the extreme — the company and culture dictate that you spend every waking hour on Tesla work¹, usually just on weekdays.

• Impact: Like undergraduate, you and your coworkers go through trauma bonding. This also means there are plenty of horror stories to share at dinner time. Most importantly, you completely disappear from your life outside of work on weekdays — social circles, friendships, even family. Quite simply, if you aren't showering, eating, commuting, or sleeping, you're working.
• Comparison: No other company I've worked at even comes close to this, for better or for worse — not even the early-stage startups I've worked at, REX and Deepscale. Effectively, any other company's work-life balance is be better than this. Granted, as a graduate student, I was certainly no stranger to working all the time. However, time pressure made constant work into constant stress.

Con #2. There is no job security. At a high level, the more exposure you have to Elon, the more likely you are to be fired. Even without E-exposure, the lowest performing members of each team are periodically let go.

• Impact: As an intern, I was fortunately not impacted by this. However, this stress had a strong adverse impact on some of my coworkers, several of whom eventually left both during or right after my internship. Shortly after, several others were then laid off. The constant churn keeps you on your toes — either you, or your closest coworkers may be laid off at any time. In turn, this makes it very easy to burnout.
• Comparison: Again, no other company I've worked at even comes to close this. Based on hearsay, Amazon may have the closest job instability. For example, during my several years at Meta, no one was let go, and no one left the team. This certainly has changed with recent layoffs, but even that was considered an anomaly. Tesla's churn is no surprise.

Con #3. The job is unpredictable. In short, direction can be dictated and changed by Elon in a moment's notice. This makes it difficult to plan too far in advance, which is partly why the team plans for the near-term only.

• Impact: Most importantly, this makes the work stressful. If Elon has a bad drive — e.g., he needed to intervene when full-self driving was engaged — he will email blast the entire AutoPilot team. During my first week, we stayed at the office until midnight, as the full-time scientists presented work to Elon one-by-one. Fortunately, as the intern, I was spared from presenting.
• Comparison: Effectively, outside of product launches and paper deadlines, there are no reasons for late nights at Meta, Apple, and others. Even then, those are pre-planned dates, set well in advance. Nothing compares to the unpredictability of work stress at Tesla.

As a result of all the above, staying at Tesla AutoPilot for 1 year is considered a huge milestone. A good half of the scientists had been at Tesla Autopilot for less than a year. Another quarter for less than 2 years.

Join for the work, not for the workplace.

With 20-20 hindsight, I now would say that Tesla Autopilot is worth joining for the work and the team — not necessarily the workplace. The culture is ideal for anyone looking to learn how to hustle. Quite simply, the culture brings out the most productive you, you've ever seen.

I'm really happy I got to experience the Tesla AutoPilot working environment, but I'm also glad that experience was limited to an internship. I don't believe this stress is sustainable in the long-term, certainly not for myself as a full-timer.

The long hours are not the problem so much; even at my current job with Apple, I'm more than happy spending every waking hour on work. However, the unpredictability and lack of job security make those hours unnecessarily stressful. Furthermore, the expectation of around-the-work clock is too much: If I'm burning out, there's no outlet. At any other job, I can simply scale back to a normal 8-hour workday for a week or so. Take a few vacation days to relax.

I also would want around-the-clock work to be recognized as above and beyond. This is for a simple reason: If I can spend 14 hours working every day, I would rather spend those hours at a team that isn't expecting this at all. Rather than putting in exceptional work just to keep up, I'd rather it exceed expectations.

In short, I'm really happy I worked at Tesla Autopilot — for the team and the work, but the stress is a big challenge for long-term careers at the company. If you do join, maximize your time there by learning all you can about how your team is so productive.

Why I worked at Tesla

Here's a brief addendum summarizing my own experiences at Tesla. In late 2020, I wasn't actually too interesting in pursuing an internship. However, recruiters from Amazon, Nvidia, and Tesla reached out, so I decided to interview and learn more.

In February 2021, I was committed to interning that summer. It had been 11 months since the start of COVID lockdown, and I was looking for an internship to kick me out of my lazy stupor. Being at home for nearly a year, for the third year of my PhD, had really sucked the joy out of the program.

For these reasons, I was looking specifically for an in-person internship. This ended up being a very effective criteria for deciding where to intern: Amazon and Nvidia were remote, without a certain future date for in-person work. Tesla was fully in-person. I committed to Tesla.

I loved the team at Tesla, and I really do think the internship opened my eyes to how productive I can be — when I'm focused and grinding. However, when searching for a full-time job in 2022, I simply decided I wanted to focus on computer vision applications in augmented reality instead of self-driving cars.

← back to Guide to the Job Hunt