Understanding Tesla FSD Beta v.10.11 Improvements w/ James Douma

[@bacontator6855]

It's a real treat every time we get to listen to James and yourself break down the updates. Especially since we got to watch the entire deployment of each step of Tesla FSD through your channel.

[@professortweedypooper3549]

My route has a right turn followed by a left turn as James described. 10.10 would turn right, into the right lane and struggle to get into the left lane in time to make the left turn. Today with 10.11, it made the right turn directly into the left lane very smoothly Much improved.

[@ken830]

Any video with James Douma is an instant thumbs-up even before watching... guaranteed to be a densely-informative watch. Absolutely great content. Thank you sooooooo much for sharing!

[@ken830]

Okay.. now that I've watched it... I feel bad I can't give it a second thumbs-up...

[@BoeingPrototype]

I am just a night time wage slave and these videos really help see me through till the bitter end. I appreciate this. Thank you James and Dave.

[@marcel5125]

We all gonna make it, bro

[@StephBougie]

Nothing better than coming back from work and find a new Dave/James episode to watch while having a cold beer!

[@polishfish]

James is a superstar! I hope he never stops coming onto your channel, Dave!

[@ev4k916]

Always great to get James' simpler explanation of what improvements he sees! Great video guys 👍🏻⚡

[@lorenblaine5275]

Appreciated having James be able to spend some time thinking about and going into details, vs making content that is easily digestable for the largest number of viewers.

[@bobdyck8508]

I really enjoy when you and James get together. My understanding of the FSD is growing as the FSD develops. Thank you so much.

[@TheLastMoccasin]

Could not agree more with James and having testers that have "skin in the game"!! I think it's the safest way of going about it!

[@casvanmarcel]

I am so grateful for having the opportunity to listen to these talks between you two guys. Thank you from the heart!

[@iselledesignstudio]

1.5 hours of pure goodness FSD technology update from James ❤️

[@endeuinable]

If it wasn't for this series of interviews with James, I would have no clue about how FSD works at all. It's a fascinating topic and James does an amazing job at breaking it down interestingly and in clear terms.

[@FailTrainS]

Thanks James and Dave for putting your time into these analyses! I always appreciate the perspective and unique insights and opinions that you provide.

[@kstaxman2]

I could listen to you two for hours on end. James of course is such a treasure as he can help us understand what is so very difficult to grasp, but it's more that than. You are really great at helping guide people with your questions. You just seem to know when to interject a question without asking too many or making the questions take too big a step or get off track. That's an art few do as well as you do. So together you two make a team that I don't think there are many out there that do a better job and make it so enjoyable it's easy to watch for an hour and a half. Just keep up the good work.

[@JoelEZ1]

Wow, I've learned so much of AI in the last year. I actually was also looking at 10.11 vids and was also looking at the visualization compared to the actual intersections. Glad I'm connecting the dots (no pun intended) and seeing a few of the same things James is looking at for improvement metrics. Thanks to you guys for providing the Tesla reports and the translation to the masses!

[@bernhardkaltenboeck9894]

Thank God a Douma interview to brighten our mood!

[@rdyer8764]

Yup! We LOVE James! Definitely my favorite series of all your interviews. :))

[@ndrewelliott8022]

I love watching these two gentlemen...great discussion

[@rodneysmith1750]

Wow, another fireside chat with James and Dave, love it! I get so much value from James's description and evaluation methodologies that I just don't get from many others on youtube. Thanks, guys for another (#28) excellent, informative, revealing, and enjoyable chat!

[@roddlez]

Thank you, James and Dave! Nerding out about Tesla's AI for an hour and half, yes! These are my favorite episodes!

[@cwmaguire]

I really enjoy these conversations with James Douma. Very informative. I like how he stops himself when he senses that he's said something people might not understand, and breaks down terms and ideas.

[@sangeetbrar7290]

I was thinking of asking when you were going to talk to James next!!!! Thank you both

[@mhfs61]

Great lecture guys. Got educated. Really enjoy James’ enthusiasm. Thank you Dave and James.

[@amosjoannides]

Love it. You know, another reason u-turns might not be in the cards yet is the turning radius. My Model Y definitely doesn’t tuen as sharply as any vehicle I’ve ever owned before. So you know, maybe

[@johnmqueripel2367]

Brilliant as always, James is incredibly good at explaining complex systems.

[@fiddlerJohn]

Love James Douma. Thanks Dave.

[@Fedzy]

Man I love these chats with James - thanks guys!

[@mackxue779]

Another great conversation! Insights and truths are so hard to come by given all the noise. Thank you for these perspectives!

[@asafprivman]

I've seen every video you released in the last 6 months, this is by far the most informative and important video thus far about Tesla

[@Bartender_Bill]

SO LOVE THESE INTERVIEWS WITH JAMES!!!

[@DonBurke1]

Every one of these discussions with James is just the best. While watching this one, I was thinking how cool it would be if someone videotaped Tesla's FSD team watching this discussion and capturing their reactions to James' observations about what is happening with the NN or what the team is doing to move things forward. What would their reactions be? Maybe "Wow, does he understand what we're doing so well?" or maybe "If he only knew what we were really doing." or maybe "He is totally off base." I really hope there is someone documenting the progression of FSD from the inside so we can all get some insight into how it really went down at some point in the future.

[@DrewS777]

James is amazing your best. So interesting. Dave you really put some Great guys in our streams of consciousness. Douma and Emmett are two

[@electrified0]

As someone on the beta, the previous lane calculation solution is by far the biggest weakness leading to my disengagements and interventions so I'm glad to see it being reworked. It struggles with the intersections leading in and out of my neighborhood, since the turn lanes are unmarked and often turn into parking lanes. Hopefully this change allows it to learn from my disengagements and I can finally start getting some zero-disengagement trips.

[@jimmyjin7200]

I really enjoy when you and James get together. Everytime when Dave Lee post on the KZbin, I must watch it, Dave Lee your post was the best I have even seen, My understanding of the FSD is growing as the FSD develops. I think with Eln Musk very intellegent reserch team, they will finsh FSD all programm by end of 2022 year, Thank you so much for your work and post !!!

[@gressex7]

This is so awesome. Thanks Dave for being James to us!

[@tstan6827]

Fantastic video! I have been in the FSD Beta since 11/20/2021, and this video has taught me a ton. Can't wait until the next major update.

[@edwardtse8631]

James talking about complex topic like FSD is like singing a song to me, it's both soothing and simple to get its deepest meaning..

[@orangektstreet]

Thank you for James and Dave for these Tesla FSD interviews. So much insight and knowledge :)

[@spacegirl2605]

Great show as always - always interesting when you have James on! I learn more every time!

[@flyshacker]

1:23:30 - Really interesting point by James that the navigational routing should take autopilot’s abilities into account, and take difficult stop signs on busy streets versus easier traffic lights into account when plotting a route. I can also imagine this as an autopilot setting for the driver: (1) Plan shortest route; (2) Plan route that is easiest to navigate.

[@ronaldlenz5745]

The explanation of proper recognition when close to big trucks, when in AP was very helpful. Going eastbound on I-10, east of Tucson, the road gets hilly/twisty and there are places with little to no shoulder. One time AP almost ran me off the road because it thought I was too close to a truck. Now I turn off AP if the condition looks sketchy. Can't wait for v11 to solve some of these AP problems.

[@rich1383yt]

1:23:20 I believe UPS GPS routing code heavily favors right turns over left turns for route optimalization.

[@grlldfsh123]

This one got me excited! The 10.11 update was significant but confusing and James really makes it clear here. What a wonderful conversation. Thanks! When is the interview with Elon? I know he agreed to one and invited you!

[@aarondawe8640]

James and Dave! What a combo!

[@josephlemieux5909]

Thanks Dave excellent interview with James.

[@NinjaAndRib]

Love these episodes!! James is amazing 🤩

[@basembarakat]

As for the question why Transformer is done now and not before? In my view Transformers in vision network is relatively new. These were first showed up in speech and natural language processing such as GPT as mentioned by Andrej. But recently researchers started suggesting to use them in vision networks. In my view Tesla is probably among the first to utilize them in real product!

[@9229kimhong]

Very delicious conversation 😜😋 Thank you.

[@edlam8325]

James always give great insights.

[@CorwynGC]

I have been looking forward to hearing James thoughts on this ever since I saw the 'James Douma fix' in the release notes.

[@rb-qc1fn]

Another gift. Thank you Dave and James.

[@codemark7464]

excellent explanation on transformers, thumbs up to James

[@GreylanderTV]

56:35 I go through exactly the same thing. Sometimes I get this let down feeling discovering they are using some approach that is incredibly simple or "brute force", and then they'll describe something that is just "wow, now that is cool".

[@colinmackie5211]

Excellent discussion. Neutral networks are central to AI but such a "black box". Please keep the discussions going. I'd like to understand more about how "training"works. I'd also like to know how that prevent changes from accidental undoing a well trained network. Can they and do they, back up a neutral network?

[@Harryandleo]

Great video Dave and James

[@ronking8726]

Thanks. Wonderful insight into the improvements and limitations of the software.

[@PhilippeLarcher]

31:30 that's the vietnamese approach, there are lane going in and out of the intersection but anything inside the intersection is another world

[@stefanherzberg1630]

Always fun and informative, thanks!

[@johnmacri7440]

Great insights. Love hearing both of your thoughts

[@stevedowler2366]

The other day, in my non-self-driving (NSD) car, I was at a stop light in the go-straight-ahead lane with left turners on my left. The first left turner was big and tall so I couldn't see over him on my left. The light turned green for both of us (left turners and straight-headers) and I started to move but the big left-turner didn't. Then a kid came running across on the crosswalk, catching up with their friends but crossing against the red light right in front of me. I saw that and hit the brakes in time but I'm not sure an FSD car would spot that in time. Or actually it would see it but really jam on the brakes hard where I applied them only as much as it took to stop in time.

[@williamwoo866]

Thanks Dave and James

[@Delcredere100]

Very educational. Thank you both.

[@ericq9844]

another great interview. thanks both. QUESTION: do you think Tesla can use huge database and NN to draw high definition map and use it in their FSD? The high definition map can be continuedly enhanced by additional data. the high definition map make all landmarks handy for FSD.

[@optimagroup11]

You gotta love James Douma!

[@robinwang1129]

Hi, Dave, like your channel, very insightful and inspiring. Just scary to see video longer than 20 minutes. LoL 😆

[@thek3743]

Super interesting, thank you both!

[@MikeCasey311]

Fascinating although I understand little. Is the neural network located in Fremont or in the car?

[@goingballisticmotion5455]

I rented a car in Europe that displayed the speed limit and it was constantly wrong (a French brand). Like James I tried to guess what it was doing. It seems the speed limit signs are GPS tagged, so when you pass one it changes the speed limit. But it seems to look for signs within a radius of the vehicle. Hence, when on the highway, if you pass under a bridge, and if that road has a speed limit sign nearby, suddenly you get that speed limit for the highway. This is an example of poor execution and relying on map data, with on context of the road the sign belongs to, and the road we are driving on. It was obviously because we would get speed limits like 30 kph on the German Autobahn, which is not a valid speed limit. Always after passing under or over a bridge. But not all bridges.

[@mfpears]

When I first learned what transformers were, I thought they were basically performance optimizations for RNNs. It's basically the same. It allows you to interrelated neurons as if in a sequence. I bet some problems will still need RNNs though

[@cathyk9197]

James! When you get FSD Beta 10.11, please drive down to Irvine and see if FSD can achieve left turns if 2 L turn lanes have a choice of 3 lanes. Would try myself, except don't have FSD Beta yet on new Tesla.

[@erlin7878]

Thanks for all your insightful interviews dave. May i request u to also interview Chicken genius singapore from time to time? I watched both of ur interviews with him in episode 173 and 223 and its been a very long time so would be grateful to have both fav utuber of mine to give us ur insights. Thanks

[@fargom54]

Some comments: 1) Agile: why NN vs bag of points, now: small deliverables over the entire end product. 2) Q: if NN has lane intelligence, how is that stored in your vehicle, since we're not always connected to the network? 3) why do we think NN can understand intersections humans have issues with? Every day I lean in the driver's seat trying to find the lane OR the correct lane. Thanks. Love the interviews!!

[@CorwynGC]

1) Not sure what you are asking. 2) The NN is stored and running on the hardware in your car. It isn't a network in same sense as the Internet. 3) The NN has the advantage of being taught by every human that has gone through that or a similar intersection (in a Tesla). You only have your personal experience.

[@bertwright1790]

My experience with phantom breaking and “tar marks” relate to the autopilot stack, not the FSD stack. So does FSD beta 10.11 have fixes for both stacks?

[@Global_Optimization]

No, Beta V11 will be the merging of the two stacks

[@budgetaudiophilelife-long5461]

🤗 THANKS DAVE AND JAMES…FOR MIND BLOWING 🤯 CHATS 🤗😍😍😍

[@kenschwendemann]

Great discussion, thanks for doing it

[@jeanderbar]

So we have now got rid of the C heuristic that used a 'bag of points' from the perception system on lane markings to designate everything, where it's now a NN. This is then fed to the planner. Is the planner itself hand-coded in C or a NN?

[@rogermarnoch767]

Wish you had a show with James every week.

[@bluetoad2668]

James Douma overload, too much of a good thing.

[@NickMackenzieMD]

Excellent as always!

[@HeavenGuy]

James is one of my favs.

[@phildavis9671]

Why can’t the maps be updated with the information for Tesla’s visualization?

[@redcrumb]

woah thanks so much for the education James!!!

[@julianriise5618]

If there is an audio auto-level in your editing software, please apply it, James' voice is relatively low. Thank you, great content!

[@DanielZajic]

This is so good, thank you both!

[@georgeholloway3981]

This is absolutely amazing!

[@pazznecht]

Thank you Dave. :-)

[@beanlovelife]

Can we take the amount of total return by sale that number amount?cause next day I see them gone sorry new invest here please help

[@christopher_rubicam]

I would love to hear George Hotz views about CommaAI vs Tesla FSD development.

[@leesweehuat]

Wondering if Tesla FSD can have an indicator on the confidence or danger level of the FSD, so as to let FSD users when they have to pay more attention or have more interventions. E.g. when vehicle at expressways, the owner vehicle colour is green, crowded city streets, owner vehicle colour is red, own vehicle colour is orange in moderate driving situations. With 3 shades of green, orange and red, 9 different levels of FSD confidence or danger level can actually be shown. Maybe this can be done in 6-12 months time?

[@ruhdisthefirst8279]

Thanks Dave!

[@gridcoregilry666]

this guy is pure knowledge, thx Douma! You should apply at Tesla AI

[@TheElectricMan]

thanks for the inteview

[@carensimon4236]

Dave - 2nd time I've seen James - an introduction for him would be nice next time.

[@Steve_0_1]

Looking forward to when Teslas can do U turns on FSD. I do one every day on the school run. Thanks James, as always for your great insight and of course thanks Dave for all your videos, I watch them all. I like to donate to charities that you support like the Afghanistan one recently because I know if you are supporting it then the money will go to the people who need it the most.

[@HeavenGuy]

Why not add a confidence button with a sexy button like feature injecting can data feedback. When the neural net makes a decision but then hesitates, press the correct decision button to increase the network confidence (train it more directly at the driver level).

[@michaeld7373]

Great job again

[@chantalpique]

Lately, I found too many fake KZbinrs are destroying Tesla! Somebody got to stop it! Your interviews always comfort me with insider Tesla info from James and many of your intelligent selections! But, what about the discussion of one of the Neuralink engineers?

[@teslahype]

The graph paper analogy 👌

[@rmkep]

Interesting discussion Dave/James. My anecdotal data suggests that many people (mostly males) like to drive a vehicle regardless of their ability to do so safely. My interest in FSD is minimal and other characteristics of Tesla vehicles is the draw. My wife who is a similar age (over 60), is looking forward to a fully functional FSD allowing her vehicle drive her (when I'm not available). We both drive Teslas and love the cars as they are, with the exception noted above. We have no interest in beta testing FSD but we are gratified that many Tesla drivers are willing participants in developing FSD. Their contributions to this complicated development are huge and while potentially messy, using current owners FSD data is brilliant and is the only way to get FSD done in a timely manner.

[@adamrak7560]

I have tried to used transformers to run quantum chemistry on molecules. They have worked amazingly well.