As Tesla started rolling out its Full Self Beta system to Canada, Tesla CEO Elon Musk revealed that the company’s AI Director, Deep Learning expert and OpenAI founding member Andrej Karpathy, is currently on a ~4-month sabbatical. Karpathy confirmed the news on Twitter.
According to the AI Director, he is currently taking some time off Tesla to rest and travel, though he is also excited to refocus and re-sharpen his technical edge. The AI Director further noted that he is looking forward to getting back to his work at the electric vehicle maker.
“Taking some time off to rest & travel after almost 5 years at Tesla. Esp(ecially) excited to get focused time to re-sharpen my technical edge and train some neural nets! Though I already miss all the robots and GPU/Dojo clusters and looking forward to having them at my fingertips again,” Karpathy wrote on Twitter.
While the AI Director’s sabbatical may seem like worrying news to Tesla supporters considering his role in the development of Autopilot and Full Self-Driving, it is difficult to deny that the Karpathy has helped set the company’s self-driving efforts on a potential path to success.
Upon Karpathy’s return to Tesla after his sabbatical, several important and exciting projects will be waiting. These include the deployment of the company’s Dojo Supercomputer, which is expected to help accelerate the training of Autopilot and FSD, as well as the rollout of the company’s Hardware 4.0, which will be deployed in the Cybertruck.
Andrej Karpathy is considered one of the most prolific experts in the Artificial Intelligence and Deep Learning fields. During his stay at Stanford, he designed and became the primary instructor for the university’s first deep learning class, “CS 231n: Convolutional Neural Networks for Visual Recognition.” The class grew to one of the largest in Stanford, growing from 150 enrolled students in 2015 to 750 students in 2017.
Karpathy was also a founding member and Research Scientist at OpenAI, an initiative that was, at one point, linked to Elon Musk. The Tesla AI Director worked on deep learning for generative models and deep reinforcement learning in OpenAI before he was hired at Tesla as the company’s AI Director. Karpathy currently leads Tesla’s computer vision team of Autopilot, which focuses on data gathering and neural network training, among others.
The Teslarati team would appreciate hearing from you. If you have any tips, reach out to me at maria@teslarati.com or via Twitter @Writer_01001101.
Turn-by-turn navigation is not new technology.
For over two decades, drivers have relied on Garmin, TomTom, and later smartphone apps like Google Maps and Waze to receive precise, reliable directions. These systems have guided millions safely through unfamiliar cities, highways, and backroads with remarkable effectiveness. They handle real-time traffic, construction detours, and complex intersections with minimal fuss.
Yet Tesla, the company that promised revolutionary Full Self-Driving (FSD), continues to struggle with this foundational capability. As FSD (Supervised) v14.3.4 has started rolling out to cars this week, navigation remains its glaring Achilles’ heel, undermining the entire autonomous vision.
Tesla Summon got insanely good in FSD v14.3.2 — Navigation? Not so much
Tesla’s FSD excels in many driving behaviors—smooth acceleration, confident lane changes in ideal conditions, and responsive handling of visible obstacles. However, when it comes to following a route accurately, the system falters repeatedly.
Owners report wrong turns, missed exits, inefficient routing through local roads instead of highways, phantom speed limit errors, and even directing vehicles to building rear entrances. Interventions for navigation issues often outnumber those for core driving maneuvers. Tesla has begun surveying owners specifically about these errors, acknowledging the problem after years of complaints.
Navigation is perhaps my biggest complaint when it comes to FSD, because sometimes, we do know better. Some of us have been living in our areas for our entire lives, but even those who have not have years or even decades of experience driving on local roads. We might know a little better about routing.
But the navigation mistakes are more than just FSD potentially taking a slightly different route that may or may not save you a few minutes. Sometimes, they’re genuinely mind-boggling.
This isn’t just annoying; it cascades into broader failures. A flawed route plan confuses the AI’s decision-making, leading to hesitant behavior, unnecessary disengagements, or dangerous maneuvers like attempting impossible U-turns or ignoring clear ramps. In a system meant to operate with minimal supervision, unreliable navigation erodes trust.
More often than not, false or plain incorrect navigation is what causes me to interrupt FSD operation. Unfortunately, I believe the latest FSD version is the worst example of it, and it leads me to believe that Tesla might be making some changes; they’ve just made them in the wrong direction.
It makes you wonder: Why is a company that has done so much with the progress of FSD and autonomy struggling so much with navigation, something that is not new and has been around a long time?
Multiple Data Sources
First, Tesla’s navigation relies on a fragile patchwork of multiple data sources—Google Maps, TomTom, OpenStreetMap, Valhalla, and its own fleet-derived data—stitched together rather than a single authoritative map. When these conflict on lane geometry, road status, or turn details, the system hesitates or chooses incorrectly.
Traditional GPS providers maintain centralized, regularly validated databases with professional curation and rapid updates. Tesla’s hybrid approach, while innovative in crowdsourcing, introduces inconsistencies that a purely vision-based or end-to-end AI approach may not easily reconcile in real time.
Persistent Learning
FSD seems to struggle with persistent learning from driver interventions.
Unlike consumer apps that quickly adapt to repeated corrections or user preferences (e.g., avoiding certain routes or remembering habitual detours), Tesla’s FSD often fails to internalize fixes on the same trip or across similar scenarios. Owners note making the same manual override multiple times without the routing engine updating its behavior meaningfully.
This stems from the neural architecture prioritizing real-time perception and control over long-term route memory and personalization, making navigation feel rigid and “opinionated” compared to the adaptive logic in Waze or Google Maps.
I noticed that when I asked Grok to try and get me home a certain way (a way that FSD routinely took in the past because it was the most efficient), it had to place a waypoint between my location at the time and my house. When I went to edit the waypoint out, as Grok had placed it for a way to get FSD to get off the highway at the right exit, it was stumped again, rerouted, and took a longer way home.
The next thing I’ve noticed, and this might be controversial, is that Nav has gotten even worse.
I think that might actually be a good thing; Tesla seems to be adjusting it. They just need to adjust it the opposite way.
The car is taking extremely strange routes to very… https://t.co/UHg3tVfNA2
— TESLARATI (@Teslarati) June 16, 2026
Reasoning, Scaling, and Intuition
Third, scaling navigation for unsupervised or robotaxi ambitions requires not just accuracy but adaptability and user-like reasoning. Current FSD often defaults to single routes that ignore driver preferences or real-world nuances like time-of-day traffic patterns. It fails to match the intuitive, context-aware planning that traditional systems have refined over the years.
Resolving navigation is critical for several reasons. Practically, it is the backbone of any autonomous journey: without trustworthy routing, the car cannot reliably reach destinations, rendering FSD useless for robotaxis or hands-free commutes. Safety depends on it—mismatched plans create hesitation in merges or intersections, increasing accident risk.
Economically, Tesla’s valuation and future hinge on FSD delivering unsupervised driving; persistent navigation flaws delay regulatory approval and erode consumer confidence. For owners who paid premiums for FSD, these issues represent unfulfilled promises. While it is unlikely Tesla will lose too many customers due to bad navigation, some will be frustrated with the constant need for human input.
Tesla has achieved miracles in electric vehicles and battery tech. Mastering turn-by-turn—technology Garmin nailed in the early 2000s—should not be this hard. By investing in tighter data integration, faster learning loops from interventions, and more intuitive routing algorithms, Tesla could close this gap.
Until then, FSD’s navigation struggles highlight a humbling truth: even the most ambitious innovator must sometimes master the basics before conquering the future.
A Tesla Cybertruck driver in the United Kingdom had their all-electric pickup seized by local police in the Greater Manchester area after the department cited “legitimate concerns.”
Last Thursday, police saw the pickup on the roads and decided to pull the driver over. Greater Manchester Police said:
“Whilst this may seem trivial to some, legitimate concerns exist around the safety of other road users or pedestrians if they were involved in a collision with the Cybertruck.”
🚨 A Tesla Cybertruck, which is illegal to drive in the UK due to safety concerns, has been seized by police in Greater Manchester
“Whilst this may seem trivial to some, legitimate concerns exist around the safety of other road users or pedestrians if they were involved in a… pic.twitter.com/cqhdPok3DM
— TESLARATI (@Teslarati) June 16, 2026
The Cybertruck in question was, according to the BBC, registered and insured abroad and was confiscated. The driver, who is a UK resident, was reported.
The Greater Manchester Police Department then added:
“The Tesla Cybertruck is not road-legal in the UK and does not hold a certificate of conformity.”
The Cybertruck cannot be legally driven in the UK because it has no UK Type Approval for operation in the country. This is due to some safety concerns, which are related to its angular shape and design. The stainless steel exoskeleton has sharp edges and projections that violate UK/EU rules on pedestrian protection.
Tesla has considered creating what it referred to as an “international version” that would be approved for operation in Europe. However, there has been no real movement on that front by the company, as it has been focused on the Robotaxi rollout primarily.
A new report claims that Apple is in the process of developing what would be the missing link for Tesla to get CarPlay.
Apple and Tesla have been reportedly working together for some time to give Tesla owners the opportunity to utilize CarPlay within their vehicles. While many owners are more than happy with Tesla’s in-house UI, which is seamless, effective, and smooth, some still want CarPlay, which does have its advantages.
A report from 9to5Mac now states that a new CarPlay technology that was highlighted during the Worldwide Developers Conference (WWDC) would potentially be the bridge between Tesla and Apple. With the addition of a feature known as “Route Sharing,” which gives a navigation app the ability to share routing data with the vehicle, Tesla would be able to launch CarPlay in its vehicles, the report states.
CarPlay has not been a priority for Tesla because it has done extremely well with its in-house UI, but some drivers are just used to it. Additionally, it could improve Tesla’s subpar Navigation or offer improved app capabilities, especially with iMessage.
Route Sharing is an intended addition to CarPlay’s iteration in iOS 26.4, which was released in March:
The addition of CarPlay would undoubtedly be welcome, but at the same time, it seems like Tesla realizes it is not of the utmost priority. There are so many things that Tesla is working on currently within its own vehicles, especially attempting to solve self-driving.
Back in February, Bloomberg had reported that Tesla was still working on bringing CarPlay to its vehicles, but it had not due to app compatibility issues and incredibly low adoption rates of iOS 26.
This bottleneck could buy Tesla the proper amount of time to develop CarPlay for its vehicles. It would be a welcome addition, and could be brought on with either the Summer or Fall 2026 Software Updates.
Tesla’s Navigation Nightmare: Why the easiest part of FSD might be the hardest
Tesla Cybertruck driver gets pickup seized for ‘legitimate concerns’ in UK
