Tesla FSD v12: The Quest for General Driving Intelligence (Ep. 753)
ฝัง
- เผยแพร่เมื่อ 13 พ.ค. 2024
- I discuss why Tesla had to do a complete rewrite of FSD and more importantly, why it should allow them to improve faster and potentially solve autonomous driving.
Narrow Driving Intelligence (NDI) is an approach to autonomous driving where each unique driving challenge, or edge case, is solved individually with a specific solution or algorithm. However, this approach does not improve the vehicle's ability to handle new, unseen edge cases.
On the other hand, General Driving Intelligence (GDI) aims to develop a holistic driving intelligence that can handle a variety of situations and new edge cases, much like a human driver would. It involves creating a massive neural network that learns from millions of videos of various driving situations, enabling it to intelligently navigate through situations and edge cases it's never seen before.
Social
🐦 Twitter / heydave7
🎧 Apple Podcast: podcasts.apple.com/us/podcast...
🎙️Spotify open.spotify.com/show/2iin015...
📸 Instagram / heydave7
Dave Lee on Parenting, / @daveleeonlife7659
NOTE: Please don’t fall for scammers who might impersonate me or others in the comment section. I do NOT give out my number or ask people to contact me.
💁🏻♂️ Disclaimer: All content on this channel is for discussion and illustrative purposes only and should not be construed as professional financial advice or recommendation to buy or sell any securities. Should you need such advice, consult a licensed financial or tax advisor. All views expressed are personal opinion as of date of recording and are subject to change without responsibility to update views. No guarantee is given regarding the accuracy of information on this channel. Neither host or guests can be held responsible for any direct or incidental loss incurred by applying any of the information offered. Author is long TSLA and other stocks at time of original video publish date.
Check out my archived articles/posts on Tesla and investing: teslamotorsclub.com/tmc/threa...
#Tesla #TSLA #Stocks
![[TH]FFWS SEA Spring 2024 Knockout Stage - Day 18](http://i.ytimg.com/vi/IfsyX-nCiZQ/mqdefault.jpg)
hey Dave it’s great you’re back!
I hope everything is OK with your family 🙏🏼
This calls for an interview with James Douma!
So happy seeing Dave’s videos again
love Dave, balance and informative. Of course I do my own DD but I love his thought process.
Glad you're back Dave.
You always do a wonderful job breaking things down to a level non engineer/ software minded individuals could understand!
Beautifully explained Dave. The underlying principals that will allow Tesla to solve for General Driving Intelligence is also how they will solve for general A.I. in the Bot. FSD using this new way of understanding, learning and moving through the world with growing experience and ability to problem solve will be amazing to watch. The speed at which this will happen from THIS POINT onwards will be exciting.
This was a summary I could show to anyone who doubts the progress of FSD or Tesla's approach to solving it. Thank you.
Beautiful clear explanation Dave, thank you so much!
Really appreciate you Dave!
Dave is very true to heart with HIS VIEWS!
When is James Douma coming on wrt v12!.
This process gives me much more confidence in FSD. What must be shown to the individual is that as time goes on the unit operates better over time. Not producing the same error in every situation, but showing the operator it too is improving. We must learn from our mistakes otherwise trust by the operator will turn into a major issue.
Thanks Dave. Most helpful!
Excellent explained, thx
Dave, great video but as a $tsla bull and investor I have to criticize elon and tesla why it took them so long to discover end 2 end neutral nets is the way to go. This has been a major setback for FSD level 4 and robotaxi and I have to wonder what the impact is. This was a hugh mistake to not realize this sooner just like not releasing or working on model 2 compact car sooner. We should criticize elon and tesla when warranted in hope they learn from this and re-examine some of their current thesis or approach as no one is perfect.
exactly the point I was typing down a minute ago - why only now ?
Thing is, if you criticize Elon for anything, you are called a "hater".
How can they know it's the right way to go if it wasn't fully realized? Before Chat GPT, these were just pipe dreams. They believed/knew it could work, but they didn't know how effective it would be, and therefore to even consider extrapolating that or creating technology analogously would have been something that was very difficult to consider. Chat GPT changed that, and everyone in the world realized the potential for neural nets and learning.
Great knowledge drop. Thanks
Great video, Dave! 😊
any chance of getting James Douma in for another chat now V12 is out? would love hear his insights.
Great vid Dave
Hi Dave, God bless you & your family
Thank you for this Dave Lee!
Thanks for the info.
Musk gets lots of jabs from various social media and general media over FSD, but I appreciate he was willing to start over again and again with the FSD development. The day FSD is truly ready for the masses, it will transform so many lives, both for those who drive and those who do not. Many legacy builder CEOs are more worried about what Wall Street wants, which is basically more profit now, even if it hurts the future. This is among the key reasons why most builders cannot plan beyond a few years ahead. This is also why real change is so hard for companies like GM, Ford, and VW.
I am little concerned FSD version 12 is being given to much praise, so people think it will be released to the masses in a couple of months. It is nice to know some non-employees have gotten it, but investors need to not over hype this moment. Government regulations move slow and Musk is a popular political punching bag. I do not see FSD being released out of beta for a couple of more years at the least. I think version 12 is amazing so far and gives great hope, but lawsuit happy world we live in and the FUD spewed on a daily basis, is going to make a quick release out of beta highly unlikely.
The problem is I don’t think it will transform our lives. Young or old I think we will all be dead before fully automated driving is widely available and safe in all conditions.
Be nice to see v12 able to give blue cruise a run for its money. Handsfree on some highways and some conditions would be great.
@@singed8853Depends on how old you are. I think I can see it coming. And many probably think similarly.
The criticisms is very unfair when you consider many other automakers making similar predictions. But media like to focus on Musk only.
@@Warbay57Musk is the only one charging customers thousands of dollars for a product that is 7 years late and counting.
The question is not why Tesla rebuilt FSD, but how many more complete rebuilds will be required every 3 years to get a decent product. I am thinking 5-6 more rebuilds.
So... this is a fsd beta version 1.0 again. Isn't it?
in this new end to end architecture are there still multiple "architectures" for driving vs parking? or is it all one
Dave, answer me this, how many years has Elon been promising us Full Self Driving, only to find he's now decided the way he was going about it was all wrong and so now he's changed gears and is seeking that holy grail by using this new concept. How many more years before he decides this new approach isn't the answer either and he embarks on another method?
It's darkest before the dawn, so they say.
end to end is the only way to level 5, he should of realized it sooner
Why do you think that Elon believed the heuristics approach was the way to go up until recently. ?
I'm not at all an AI expert, but having learned about neural nets some 20 years ago, how they fit data that you don't entirely understand into a model of reality, my first thoughts when I learned about FSD attempts by Tesla way back was that this is how they were working on it.
To learn now, after all these years, they are now heading into that direction as a 'new approach', well, it seems puzzling. I'm sure I'm oversimplifying things here and would love for someone to explain to me the logic of not pursuing that route in the first place. Maybe a lack of compute was the reason ? Thanks in advance !
And thanks Dave, you're always helpful !
this is a key point. Tesla has actually always been striving for the general way, this new "narrow" NDI terminology is just the latest iteration of Elon renewing efforts to make it more general, 2-3 years ago he was using the terminology moving the specific algorithm code (i.e. NDI but before the term) "over the fence" to the machine learning side. Its intuitive because we aesthetically want the problem to be solved generally, overwise its whack-a-mole as Dave Lee says, but although the general approach is the ultimate solution, that does not tell us how close we are to achieve it sufficiently (maybe it will coincide with AGI?). One of the real hurdles that does not get mentioned much is the data recorded from the fleet--by its nature, you're not going to have lots of rich data on edge cases. Humans can parse what should be done intuitively, but machine learning needs lots of data. In the meantime there remains practical pressure to try to deal with edge cases via specific algorithms (i.e. "narrowly") and so there is an ongoing tension there between general and narrow. I don't agree that this is a real shift in FSD, its just the latest spin to renew hope that its on the verge of whatever. Don't get me wrong, its impressive.
And here I thought fsd has always been neural net like alphago…Elon talked like fsd was like alphago on autonomy day.. I feel dumb now
Don’t feel bad. Musk was obscure about it. It has had neural net aspects - just not end to end. But I disagree that end to end is the best or only way to go. We need to be able to put guard rail rules into the system for local laws AND have it learn from local situations - not just one time neural net learning every month or so in my view. It’s still uncharted territory for how to really solve it.
Elon is full of lies to keep to people engaged to sell his EVs
@@longboardfella5306well of course we always need to make sure these cars are as safe as possible and rest assured tesla will build the safest car ever, they will do whatever it takes . Teslas main priority is safety efficiency and minimalism
@@longboardfella5306 Not sure which local laws you are talking about. Most of driving to my knowledge is the same from state to state. This makes sense because you cannot expect someone driving 1000 miles to memorize different rules for each state they pass through.
@@jacobnunya808 I mean cities like Melbourne and Toronto with special rules for trams and hook turns. In the US Michigan does U turns differently. Etc
is GDI approach the one taken from Comma since the beginning? Does this mean also, that Tesla may have an advantage to solve the autonomous driving problem because of driving data gathered?
I watched a couple videos from Whole Mars Catalog and from what I have seen it seems quite good. Some over cautiousness, but few interventions. That is not even the version most people will be getting.
With the general driving intelligence how does the programmer control the system? How does it say to be extra careful with pedestrians for instance?
Please explain what is meant by ‘footage feeding the neural net’. How is the training achieved? And does each node in the net refer to thousands of pieces of footage? If so how many frames of footage constitute one node or scenario? Does the fsd software refer to a node when faced with a scenario so that it can pull on a load of probabilities based on millions of frames of footage, choose an appropriate outcome and then on to the next node? Are footage frames stored in the car to enable the calc to the next node in the net, or is the footage quantified to equal a probability?
Feeding footage to a neural network involves processing each image through the network's layers in what's called a 'forward pass'. This could be done using a Convolutional Neural Network (CNN) or a transformer model, which are architectures designed to recognize patterns in visual data.
During the forward pass, the network uses its current weights to compute values at each output node. These values represent the network's predictions, like whether to make a 'left turn' or 'right turn' in the context of driving.
After the forward pass, we calculate the 'loss' for each prediction. The loss is a measure of the difference between the network's prediction (actual output) and the correct action (ideal output). For example, if the correct action is a 'left turn', but the network predicts 'straight', there will be a higher loss.
The loss is then 'backpropagated' through the network. This means that we use the loss to make adjustments to the network's weights, starting from the last layer and moving backward to the first. The goal is to adjust the weights in such a way that, if the same image were processed again, the prediction would be closer to the correct action.
This process of forward pass, loss calculation, and backpropagation is repeated for each image in the dataset. When using video footage, which is a sequence of images, the network learns from the temporal sequence as well, which helps it understand the context over time.
Once the neural network has been thoroughly trained and fine-tuned through the process described above, it is deployed into the car for what is known as the 'inference phase.' During this phase, the network operates in real-time, analyzing live video data from the vehicle's cameras to make immediate driving decisions. The network's weights are no longer being adjusted as they were during training. There's no calculation of loss because the network is not in a learning state; it's in a state of application, using the intelligence it has developed. This inference process is the real-world application of the network's learned patterns and predictions, and it's how the car autonomously navigates its environment.
So basically 8 years of work completely scrapped and they're starting over.
Yep, FSD is in reverse mode now.😂
interesting!
True...
wait for Optimus V3
Old method would never work. Thus Elon promised it year after year.
Did Tesla change its approach after "Comma" successfully implemented this approach? I do have a Comma device, and it is amazing how well it works, and more amazed when I see almost no one has heard about it, or has one.
Comma has not successfully implemented anything yet. Their system is way worse than FSD 11.
Is a Comma device better than a Colon device?
The problem with this approach is one of not knowing how the decisions are made - this is in effect a black box Ai approach. Any novel edge case may or may not be handled correctly. Regulatory nightmare. Also Musk has spoken about the new approach needing to have LLM capabilities to READ signs, not just recognise them if seen before. This is a BARE MINIMUM. We also need to FSD to be able to communicate intelligently about its intentions and to record them so that regulators can understand. Else we will have a system ok for closed environments but unable to get approached for the real world. Black box Ai approaches in medical use are currently banned for the same reason - and driving is just as life or death as lots of medical uses. We need a combination of Heuristics And neural net Algorithms (that was the name of HAL) - Arthur C Clark was onto something!
HAL just precedes IBM in the Roman alphabet.
Neural net weights act as essentially complex probability functions fine tuned via training to minimize mismatches with known reality. Since NN “reasoning” is a probabilistic black box, edge case mismatches (with what should have been done by an experienced driver) would have to be reduced by many orders of magnitude before L3 driving can be conquered
I'm operating on GDI right now
This looks to be a wildly over simplified analysis. There is no way that a rules based system could have driven as it did in versions prior to 12.1.2. It would have taken too long to run through the rules & decide what to do. Instead I believe it was assigning any aspect of driving to an internal AI that seemed suitable & that one selected AI sub processor would handle things & then pass off to another AI deemed more able for the next issue & repeat. What they have now is more like a human brain in that it is continuously looking at what must be done & using all of its artificial neurons to come up with a driving course. This should create super human driving quickly as the system is self evolving, learning what happens based on its input as would a human & it does not take most humans long to go from complete novice to legal driver with only two eyes whereas this system as far more vision & faster reactions. Thanks for sharing!
Well said.
But how long are we talking until we can remove the driver controls such as the wheel and pedals? A year, 5 years, 15 years?
I purchased full self driving three years ago when Elon was saying it would be feature complete and due for wide release by the end of the year. I must admit I am loosing the faith. My car is only running version 11.1 here in New Zealand which cant do much at all by itself. In fact I stopped using it because it was too stressful. I hope it gets released soon to the people who have paid for the sake of Tesla survival.
Did anyone notice that FSD is still in beta? FSD came out in 2016 and eight years years later it is still in beta. Musk is laughing all the way to the bank.
OK Dave, The promise of FSD which gives us robo taxi, which according to Musk is worth trillions of dollars has been promised for years, and has yet to arrive. In your view Dave, from this point forward what is a reasonable time to wait for FSD (that works) to show up?
❤
I agree. Kind of sucks to think they may have wasted some time not using this strategy sooner. All data is useful nonetheless, so its a wash i suppose. My gut is telling me this strategy will be the right one. How long until FSD? I've no clue. I do believe it will happen though
since they had to rewrite it and are coming out with v12 now, and customers have gotten little value out of it so far, will they do the right thing and allow customers to transfer their FSD purchase to new car? also will they turn down the driver monitoring sensitivity it is way too aggressive
Little value? Says you. Value is subjective.
@@CyberCPAOHobjective facts are that it is still unfinished BETA software and that it is still unable to drive car by itself
objective fact is that there was no driver monitoring when I bought FSD and I never gave permission to their system to watch and punish me like an animal, this is totally unacceptable
@@boumarc No shit Sherlock
This is like saying Google Search provided no value in its early days when it wasn’t as good as it is today. Obviously it did and FSD Beta already provides value today in its current state.
@@boumarc Cope
I suspect regulatory bodies will want some studies on how FSD is accomplished and determine some guardrails. It's unlikely governments will simply leave the solution freely up to the manufacturer without controls or oversight. I see some struggles ahead with various government agencies trying to establish standards that can be applied and measured ensuring the outcomes are as safe as possible. AI even according to Elon Musk strongly encourages governments to regulate and set AI controls.
Ooohhhhhhh 😮
The proverbial Learning Curve...
Thanks for the video, Dave. Why is Tesla using Narrow Driving Intelligence in the first place? Surely, they are aware of its limitations. Now that they are using General Driving Intelligence, it seems like a very logical choice. What has changed? Did they develop some new technology that made GDI previously impossible, and now possible?
They were using NDI because, up until AI, there was no other way to do GDI, because there was no machine/computer or algorithms that could house general driving intelligence and also make decisions. To program algorithms to do this would be beyond time and labor intensive. They were on their way to do this, but it was super inefficient and near impossible.
It's amazing that you bought your TSLA shares at $6. If you had no shares would you buy now at around $190?
Dave you finally got it, I’ve been saying all along that Tesla approach you call narrow AI was never going to work,there just too many edge cases and the time to solve each one of them was not scalable. I wish that you had more forethought and seen this before transformers were well known.
Thought that was what they were doing from the beginning, but alot was hardcoded it seems
I hope V12 is a lot better than V11 - it sucked.
I still have doubts it will correct the errors in V11 but I'll give it a try when I get mine.
Now Dave you seen the light, do you think General AI will finally get us to robitaxi? I don’t think so, there still going to be edge cases that General AI won’t be solved easily, as videos won’t be available because of the rarity of the edge case like let’s say another car hits a pedestrian ends up under the robotaxi and the robotaxi moves forward and pulls to the side of the road… now do I have you attention how is this going to be solved with General AI?
FSD is a great overall goal but we still need those everyday short term goals to supplement the business while they develop fsd. Where the flagship roadster 2, where’s the model 2 and where’s the van platform? Why is Mexico and India been intentionally slowed down? It seems from musk tweets that maybe he’s waiting until he gets a new pay package to ramp up, rather that just doing it now asap as planned. I hope I’m wrong because many people are rooting for that tesla company for him to drag his feet intentionally but one never knows for sure unless they’re in Elons head. Fed announced a pivot some would expect tesla to keep the ramp up going but who knows what they do.
The great thing about using neural network training is that Tesla can leverage its enormous data advantages to feed into machine learning. I dont think anybody will be able to get their hands on the sheer volume of data needed to make this system work. At least not for many years...
Sure FSD will change the industry. But I can hardly imagine that the impact would be that big. But the real impact could be the new neuronet for the robots. There may be the much bigger leap forward in productivity.
I had a Tesla Model 3 loaner for one day and one night as my Model X needed to have 3 things to be fixed. Tried using the FSD of the loaner Model 3 to get me home from El Cerrito supercharging location. Once engaged, the car hesitated to go forward and constantly turned to the right and left, then right and left again as if the cameras alone couldn't help to decide how far the car was from the curb on the right side. Once it started to proceed forward finally (after about 15 seconds), it turned too fast while turning to the right and the right rear wheel ran over the side curb. That was so embarrassing as the couple directly on the opposite side who were also supercharging their model Y was looking at me thinking I was driving the Model 3 and causing the tire going over the curb. Oh boy. The Tesla FSD is not ready yet no matter how many times they change the name, including this 12.1.2, no matter if they will re-name it to 12.1.3 to 12.1.1000 to 13.1.2000, it will still have problems. Forget it. My friend who paid for it for $15,000 in 2021 remember that Elon said FSD would be available by end of 2020, then end of 2021, then 2022, then 2023, now we are in 2024. It is either a SCAM or broken promise.
Just when I was thinking that the Tesla fans (I do qualify) had given up on FSD and were saying 'look at this, look at this, look at the Tesla bot' Tesla comes out with public version FSD 12. Now this comes out a few days before the Q4 results shareholder meeting and it could be that the numbers won't be that good and Musk is trying to affect the narrative, this is quite possible imo. It could also be that Musk is giving a buying opportunity to the retail market, since the institutions won't react to the FSD 12 much. (Cathy excepted, but she's already loaded up).
hi Dave. how are you. we miss you
FSD needs continual updates especially for each Country if it is supposed to be a Global product not just the US. The cross over point is maybe now that enough input has been given to the Neural Net that means the experience is actually superior to the actual human driver? Will authorities be able to accept this and does it need to be proved first??Will there be versions for each country?? Off road driving is outside the remit at the moment??
All excellent points. And a neural net cannot explain itself - that is a regulatory nightmare. Why would they trust it to not go wrong way up a ramp in a new city?
As long as it happens. And it will who really cares how long it takes..
if there is an edge case for parking at home does it throw up problems with what you know as known facts to assemble a judgement…
#hedgecase
FSD does not work even at level 3. Time to sell!
wait, the files are in the computer?
What??? Lmfao what does this even mean
A concept used by railroad engineers on extreme grades with wide variations of weight and length at different locations within their train length.
FSD starts to make sense - it evolves like Special Relativity -> General Relativity
how funny would it be that one day FSD does come into fruition but it would accomplished by another company😁
Bit repetitive. Got the concept quickly but how the new one works ?
After NDI and GDI, is the final step SDI (Super Driving Intelligence)?
I'm curious how the robotaxi will handle "control" interactions with humans outside the vehicle, whether they be security guards, police, construction workers, etc. Entering secure areas, following hand/voice signals, and dealing with a random person saying/doing random things that affect the car's safe path. All of these will benefit from GDI, but I see a need for the AI to converse/understand commands coming from outside the vehicle such that a human driver would. Perhaps there is another method to deal with the situations, but I can't think of it. Now I'm thinking about KITT on Knight Rider, ............of course I am..... th-cam.com/video/dANY3uk7lxc/w-d-xo.html
Same logic applies to Optimus. Train it to load part A into machine B, C into D, E into F etc. If it then works out for itself how to load part X into machine Y, you’re winning.
Once it learns to put D into P, it's over! Optimus will learn to manufacture itself and Tesla can get close all Optimus factories. The profit margin will be infinite!
I understand FSD 12 immediately when i figured out how ai bewbs were made 😂
FYI, here is MED BED on TH-cam
Just the best video ever. A clear concise explanation of narrow artificial intelligence vs general artificial intelligence and how Elon and Tesla are just right again in their approach to solving AGI for driving. Robots next. Go Tesla!
Fate loves irony, FSD 13 will be the one.
Because it didn’t work.
Will it work now? Hopefully. If it doesn't what else can they try?
I’m still an optimist, but as a realist, the many quirks/variables in roads, vehicles and drivers will certainly make the curve slow. Also, federal/state regulators will have to also levy unprecedented mandates on driving and data recording as it relates to liabilities, such as requiring 360 degree cameras, vehicle metadata, weather/road condition recording, etc. everywhere possible for litigating accidents. Imagine the finger pointing that will go on. I shudder to think. Regardless, FSD will eventually win.
So all the 2hr videos you did with James Douma explaining how the previous Tesla FSD was going to be successful was a total waste of time
Hello! They are the experts!🤣
Yep
True
I think AI Inferences from Audio-Video will escalate when DOJO ramps in October 2024 ! ?
I love my FSD BETA. It’s BETA. Do you know what BETA means ?
They rebuilt FSD because it doesn't work. Now it is Fully Supervised Driving v12 after years of empty promises.
What do you think of combining narrow and broad AI Like a left brain right brain set up where the left brain is rules-based and the right brain is experience-based. Let's face it sometimes we want the car to just follow the rules
“left brain/right brain” assumptions have been proven false by functional Magnetic Resonance Imaging. Those assumptions were never considered reliable by scientists. The scam did sell lots of books…
Hi, Davi. As far as I know, Tesla FSD has been using image recognition algorithm (e.g. mainly RESNET) among other deep learning algorithms since the very beginning. I don't know why you make this sound like this is new to FSD v12. I don't believe they ever relied on rule based algorithm that is widely used in industry simple automation. It seems you were misled by someone.
So the old system was vaporware but this time it will be the real deal. 12k please 😂
Yep, Elon is embracing the AI Transformer model. It's ironic as he walked away from OpenAI thinking it wasn't going to amount to much and his approach to FSD would succeed. The same goes for Optimus robots. At least he knows it was the wrong decision and willing to change course. Will Tesla be the first to reach FSD? It has some serious competition in non other than NVIDIA who started on Transformer model much earlier and recently made breakthroughs in its Foundation Agent .
All this is called VISION-BASED -AI for FSD + OPTIMUS ! ?
I do find it a strange scenario whereby folk are scared about machines driving themselves and could only ever feel safe if a human was driving instead!
Machines are possibly more than 1000 times safer then any human being could ever be and yet just that thought scares people not one little bit!
Look up the word MTBF. You will understand why FSD will not work without more sensors.
The share price is tanking.
if someone on YT can think of all these, then don't you think Tesla engineers already know these?
Just think, years from now Tezzi stock could surpass $299/share
Sell now and you can buy back 2x as many shares at 90. 299/sh may be 10 years away. TSLA is not a tech co, it's an auto, almost as boring as Ford. Investors are waking up!
time will tell how this new architecture will work out. if they can keep hallucinations at bay, that will be a big positive.
You mean LSD, not FSD!
buy the dip?
So how I understand this is that Tesla customers who purchased autopilot on potential of getting true autopilot a few years down the road on the same hardware, were instead sold on glorified lane assist heading towards technological cul-de-sac? I guess the question is at what point in the past decade Tesla/Musk knew this and decided to stick with the narrative?
No.