Article Complete
AIM
Andrej Karpathy points out a GPT-like transformer in new Tesla FSD release
Andrej Karpathy points out a GPT-like transformer in new Tesla FSD release

Andrej Karpathy points out a GPT-like transformer in new Tesla FSD release

Andrej Karpathy, the director of AI at Tesla, who also heads the Autopilot Vision team, pointed out that his favourite update in the version was that a GPT-like transformer could now predict lanes and their connectivity.

Published Updated

Key takeaways

  • Tesla's Full Self-Driving Beta version 10.11 includes a GPT-like transformer for improved lane prediction and connectivity.
  • The latest update aims to reduce unnecessary slowdowns by enhancing the accuracy of vehicle predictions during turns and merges.
  • Improvements in Vulnerable Road Users detection have been made, reducing false positives for cyclists and pedestrians significantly.
  • The update's enhancements are attributed to increased data size and modifications in training network parameters.
5 min

Tesla has released notes for the latest version of their Full Self-Driving Beta - 10.11. According to @WholeMarsCatalog, 10.11 was already in use by Tesla employees. If no bugs are detected in the version, it will be released to external users starting this week. The notes appear to be focused on cutting down unnecessary slowdowns in the model by making more accurate predictions about where other vehicles turn or merge. 

https://twitter.com/WholeMarsBlog/status/1503078975664533508?s=20&t=ZDZBD-dUdZOiQNaD4C2N7Q

Andrej Karpathy, the director of AI at Tesla, who also heads the Autopilot Vision team, pointed out that his favourite update in the version was that a GPT-like transformer could now predict lanes and their connectivity. The notes state that besides being able to predict crossing lanes, the transformer also “allows computationally cheaper and less error-prone post-processing, and paves the way for predicting many other signals and their relationships jointly and end-to-end.” 

The FSD Beta v10.11’s release notes also show that it has improved Vulnerable Road Users (VRU) detection by 44.9 per cent by reducing false positives around cyclists and pedestrians. Tesla was able to manage this by increasing the data size of the next-gen autolabeler, training network parameters that used to be frozen and modifying the network loss functions. 

The previous version, Beta 10.10, took several weeks to roll out to testers. While there is hope that this version will be released quickly, if there are any additional issues spotted, the release will be pushed. 

Membership

Stay ahead with AI insights.

Unlimited access to premium articles, exclusive analysis, and the stories shaping India's AI economy.

Subscribe now

AIM Americas

1190 Miraloma Way Suite P, Sunnyvale, CA 94085

AIM India

1st Floor, Sakti Statesman, Bellandur, Bengaluru – 560103

© 2026 AIM. All rights reserved.