Andrej Karpathy took to Twitter to announce he's leaving Tesla. The 35-year-old was the senior director of artificial intelligence and was a key part of the Autopilot team.
He did not explain why he was going; instead, he praised his team's work on Autopilot's development.
The tweet reads: It's been a great pleasure to help Tesla towards its goals over the last 5 years and a difficult decision to part ways. In that time, Autopilot graduated from lane keeping to city streets, and I look forward to seeing the exceptionally strong Autopilot team continue that momentum.
Elon Musk responded to his tweet: Thanks for everything you have done for Tesla! It has been an honor working with you.
Within minutes of these tweets, media outlets and industry observers started to speculate about the future of Tesla's FSD. Karpathy was a significant part of the team tasked with improving FSD and making it available to every Tesla owner in North America by the end of 2022.
Just 35 minutes after he tweeted he was leaving, he addressed the gossip with another tweet. It reads: Important to keep in mind that the Autopilot team is hundreds of strong engineers who very much know what they're doing, just don't have my public visibility. I was only one part of that effort, and I think get an outsized spotlight cast on me because I do.
Karpathy, a Ph.D. from Stanford University, was on a four-month sabbatical. His departure comes just days after Tesla closed its San Mateo, California office laying off 229 data annotation employees who were part of the Autopilot team. The layoffs were expected after Musk announced a hiring freeze and a 10 percent workforce cut in June.
Andrej says he doesn't have any plans for what he'll do next, but that he plans to spend additional time with his passions of AI, open source and education.
Whole Mars Catalog, among others, questioned how significant Karpathy's departure was to the company. "It's not bullish, but people are vastly overblowing it," tweeted Whole Mars Catalog. "Elon knew this would happen and that's why he started highlighting in early 2021 that Ashok (Elluswamy) was the lead on Autopilot software, not Andrej."
Nobody stays forever. Life goes on.
Elluswamy was hired as the team director of Autopilot in 2015. At the time of his hiring, Musk said he was the first person recruited for the Autopilot team.
Andrej during Tesla's AI Day
Subscribe
Subscribe to our newsletter to stay up to date on the latest Tesla news, upcoming features and software updates.
Tesla has published a 16-post series covering a wide range of unique scenarios against which the company tests each FSD release. Tesla’s VP of AI, Ashok Elluswamy, also provided some additional context, which we’ll cover below.
These scenarios could be difficult for a regular driver to respond to and are a good demonstration of FSD’s capabilities. Let’s take a look at all the different scenarios that Tesla regularly tests against. According to Ashok, these tests are only one of 10 ways Tesla validates their software. These tests were done against FSD v12.5.6.3, the HW4 build that’s on approximately 20% of the fleet.
We’ve embedded each video below and also provided some additional information.
1. Reverse Cut-in (Occluded)
Every FSD release is rigorously tested, including rare and adversarial scenarios on closed courses — Here's 16 examples:
This first test is for a car reversing out of a parking space, while occluded (vision blocked) by another car or obstacle. In the test, the Model Y notices the incoming car and then brakes with space left over.
In this video, it's hard to even spot the car on the right that’s “peeking” forward—it's occluded by bushes on the side of the road. FSD notices and stops in time to let the other vehicle safely pass.
An incoming vehicle that takes a left turn to enter the same lane of travel as FSD - spotted, stopped, and then continuing smoothly. In the second shot, you can see FSD left a considerable amount of space to let the vehicle cut in.
This is honestly one of the most common scenarios on a highway - someone cuts in while obscured into your lane of travel. Once again, FSD does its thing with plenty of space to spare.
FSD’s path planner plans an overtake around a stationary vehicle here, sees the oncoming traffic, and politely waits its turn to continue. This is another everyday scenario, especially on urban and suburban streets.
There are a surprising number of people who don’t know how stop signs work. However, FSD does know how they work and also knows what to do if others act unpredictably. This is one of the best demonstrations of FSD’s capability to react quickly and effectively to unpredictable behavior on the roads.
Another overtake scenario, but this time the vehicle being overtaken throws open its door into traffic. Not necessarily the wisest of moves, but humans are unpredictable. Good to see Tesla working to save its vehicles from both hitting the door or having the door hit with its new Blind Spot Monitoring While Parked feature.
This scenario is another common one - someone is taking a right turn, and you expect them to complete the turn, but instead, they slam on the brakes. Here, you can see the vehicle’s Forward Collision Warning goes off due to the sudden reduction of distance between the Tesla and the vehicle in front of it. FSD still manages to brake in time before Automatic Emergency Braking kicks in.
This one is pretty clean-cut - someone takes a left turn even though you’re on your way through the intersection. This particular accident is one of the most common - 53% of cross-path accidents involve a left turn through an intersection where a side impact will be lethal. FSD can make a big difference - over 8,000 people die in North America alone in this particular scenario.
Here, FSD smoothly reacts to an object on the road as the vehicle ahead dodges it. We’d love to see this scenario done with different types of objects or debris, as this is another common item - especially with roadkill on country roads. A good demo of FSD’s collision avoidance maneuvers too!
Taking it one step further, FSD successfully notices, slows down, stops itself from hitting the bouncing cone, and then makes it around safely. If you’ve ever seen someone strap something down in the bed of a pickup and forget to say “This ain’t going anywhere” - this is that exact scenario.
Another high-speed occluded cut-in, this time for a vehicle making a U-Turn from an oddly shaped intersection. FSD started braking the moment the first couple pixels of the Model Y were beyond the bush occluding vision. With a human’s reaction time, this would have resulted in a T-bone.
Ashok Elluswamy mentioned that this is possible because of low latency and high intelligence - combined together, it enables FSD to make a threat assessment and decide what to do, very quickly. It also considers the potential of a rear-end collision - and there is collision avoidance baked into that decision-making.
Similar to the cone cut-out, this one happens at 73mph (117km/h)! FSD sees the car dodging, notices there’s an obstacle, and moves out of the way without dropping speed.
Another common scenario - in fact, the number one reason why pileups occur on highways is the harsh, sudden braking of vehicles in front. FSD once again notices and comes to a safe stop here.
Tesla has been previously greatly criticized for its ability to come to a safe stop for children playing on the road. Here, Tesla is demonstrating its capability of doing so - once again with highway speeds of 73mph.
And to really make the point - it can do it with a crossing child too. Tesla has made a strong commitment to safety for vulnerable road users, and this is a good way to show how it works.
That’s a lot of unique scenarios that Tesla tests against - and each and every build has to be vetted against this huge list of different safety tests in order to ensure that the build that ends up in customer hands is as safe as possible. With FSD v13 looking like it's coming soon, we’re excited to see where this takes Tesla next.
With Musk announcing that FSD v13 is coming soon, it really is starting to feel like Unsupervised FSD and Robotaxi could be pushed out sometime next year.
Tesla announced an updated roadmap for FSD in late October. With it, they announced that FSD v13, the one that’s expected to make Unsupervised FSD feature complete (although not 100% reliable), would start being sent to employees by the end of the week.
We have yet to see any hints or sightings of employees testing FSD v13, but Elon Musk has now stated that FSD v13 is coming soon.
It sounds like Musk is talking about the employee release as it usually does, but it’s not immediately clear. Tesla originally predicted that select owners would receive FSD v13.3 around Thanksgiving.
Musk also mentioned that FSD v13 is trending to be about 500% better than the current build of v12.5.5.3 on the Cybertruck. If that’s true, Unsupervised FSD is just a regulatory leap away, rather than a technical or training challenge. However, let’s remember that Musk has often thrown out these vast improvement numbers that don’t appear to materialize when the update rolls out to customers. There’s no doubt that FSD v13 will be a big milestone and have several new features, such as Autopark at destination, Unpark and more, but the much lower miles per intervention will likely come in later revisions.
Tesla has already seen drastic improvements this year, moving from V11.4.9 to v12.3.6, and then more recently to V12.5.4.2. Each successive build has had major improvements in how FSD is able to react and respond in real-time while also becoming smoother and safer.
We’re in for some exciting times as an even better FSD version makes its way to customers soon. And in order to really make the point, Tesla has also shown off how they conduct FSD Safety Testing in a series of videos, which we’ll include below.