Uncovering Big improvements to Tesla FSD Beta w/ James Douma #38 (Ep. 693)

I'm joined by James Douma as we take a look at the most recent Tesla FSD Beta enhancements.

Andrej Karpathy interview on @Lex Fridman:

Tesla FSD Beta 10.69.3 release notes:

– Upgraded the Object Detection network to photon count video streams and retrained all criteria with the latest autolabeled datasets (with an unique focus on low presence circumstances).

Enhanced the architecture for much better accuracy and latency, greater recall of far lorries, lower speed error of crossing cars by 20%, and improved VRU precision by 20%.

– Transformed the VRU Speed network to a two-stage network, which minimized latency and enhanced crossing pedestrian speed mistake by 6%.

– Converted the Non VRU Attributes network to a two-stage network, which reduced latency, decreased inaccurate lane assignment of crossing cars by 45%, and minimized incorrect parked forecasts by 15%.

– Reformulated the autoregressive Vector Lanes grammar to enhance accuracy of lanes by 9.2%, recall of lanes by 18.7%, and recall of forks by 51.1%. Includes a full network upgrade where all components were re-trained with 3.8 x the quantity of information.

– Included a brand-new "roadway markings" module to the Vector Lanes neural network which enhances lane geography error at crossways by 38.9%.

– Upgraded the Occupancy Network to line up with roadway surface rather of ego for improved detection stability and improved recall at hill crest.

– Minimized runtime of prospect trajectory generation by roughly 80% and enhanced smoothness by distilling an expensive trajectory optimization treatment into a lightweight coordinator neural network.

– Improved decision producing brief deadline lane modifications around gores by richer modeling of the trade-off in between going off-route vs trajectory needed to drive through the gore area

– Decreased incorrect slowdowns for pedestrians near crosswalk by using a much better design for the kinematics of the pedestrian

– Included control for more exact object geometry as detected by general tenancy network.

– Improved control for cars eliminating of our preferred path by better modeling of their turning/ lateral maneuvers thus avoiding abnormal slowdowns

– Improved longitudinal control while offsetting around fixed challenges by browsing over feasible automobile motion profiles

– Improved longitudinal control smoothness for in-lane cars during high relative velocity scenarios by likewise considering relative acceleration in the trajectory optimization

– Decreased finest case things photon-to-control system latency by 26% through adaptive planner scheduling, restructuring of trajectory selection, and parallelizing understanding compute. This enables us to make quicker choices and enhances reaction time.

Apple Podcast:
Dave Lee on Parenting,.

KEEP IN MIND: Please don't fall for scammers who might impersonate me or others in the comment section. I do NOT provide my number or ask people to contact me.

♂ Disclaimer: All material on this channel is for discussion and illustrative functions only and should not be construed as professional monetary suggestions or suggestion to purchase or offer any securities. Must you require such recommendations, speak with a certified monetary or tax consultant. All views revealed are personal opinion as of date of recording and are subject to alter without responsibility to upgrade views. No warranty is given concerning the accuracy of information on this channel. Neither host or visitors can be held responsible for any direct or incidental loss sustained by using any of the information used. Author is long TSLA and other stocks at time of initial video release date. When you purchase through links in this video description, author might earn an affiliate commission.

Open an brokerage account at Interactive Brokers:.

Check out my archived articles/posts on Tesla and investing:.

#Tesla #TSLA #Stocks.


Wealth Builders Club
Wealth Builders Club Secrets Revealed – Click Here to Discover the #1 Investment Resource!

You May Also Like

About the Author: Richard Money


  1. This as got to be the best content on the tube these days. Also, my drives are nearly always intervention free now. so much fun.

  2. Just a thought, in my mind there is potentially a synergy between the exponentially improving AI and the human condition. As Andre commented AI sees reality in a different way and it’s the difference that makes a difference.🤔IMO

  3. James as always brings his special understanding to the topic. Love how he brings things together and makes us understand all the many things that are behind the scenes. And you (Dave) always bring out the best questions to stir his thoughts.

  4. For Karpathy, he got a huge golden parachute, probably more than 100M$ in bank.
    A lot of people at that point will stop working like crazy and will go enjoy themselves for a few years before working again if ever.

    1. Yheee colect art and travel, not that I personal is into travel or expencive art, but you cant blame them if they got $100M.

  5. I’ll bet dollars to donuts full FSD is still 5 or more years away. To brand it today as Full Self Driving does seem a bit of a stretch, bordering on falsehood.

  6. Two pros. Love James he is so informative and Dave pumps out quality videos like no one else. So appreciate you gents

  7. Adding to James’ point that people with less than good vision can still drive cars . . . I was surprised when my 87 year old mother lost the sight in her right eye, yet was still legally able to drive in California.

  8. Another episode that knocks it out of the park! I enjoy these so much that I usually watch them multiple times. Thanks to both of you for preparing such information packed dialogs. They certainly resolve so much more detail than I would glean out on my own. 😊😊😊

  9. For FSD summon and reverse summon to be fully realized it needs to allow for real world street, parking lots, parking structure inclines. The beta assumes that the path for summon is almost flat and stops the car from moving either forward or backward when the car detect that it is inclined more than a gentle driveway slope. Off road summon/reverse summon will be something to behold when this is fully functional 😁

  10. Hi dave, the left hand turn situation you’ve described is a debate between my friends. I was taught to turn into the second “middle”lane not the 3rd. We debate it often, my current opinion is 1st left lane turns into most left lane. Second left lane turns into middle. But some of my friends think they have the right of way into the far right lane. I disagree. As humans we cant agree.

  11. Thanks for helping to unpack the 10.69.3 release notes!

    I’m eager to see whether some of the improvements, such as aligning the occupancy network with the road surface, improve FSD’s performance on curvy, banked mountain roads.

Leave a Reply

Your email address will not be published. Required fields are marked *