Tesla FSD v12 Rollout - FIRST THOUGHTS! (Ep. 752)
ฝัง
- เผยแพร่เมื่อ 21 ม.ค. 2024
- I discuss Tesla's Full Self-Driving (FSD) Version 12, a significant rewrite of their driving software, transitioning from rule-based coding to a neural network approach. This new version is a big leap, moving from using specific rules to an end-to-end neural net. I highlight the accomplishments but also note the challenges ahead for the Tesla FSD team.
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
Good seeing you more, and I hope your wife is doing better. I am also struggling with prostate cancer. Hopefully in Mar 2024 the radiation treatment at UCSF will destroy the remaining killer cancer.
Good luck. After your recovery look into much lower rates of cancer among vegans.
Good luck! 🙏
Wish you the best of luck
My father has medicine that'll help shrink the onset of prostrate enlargement
I wish you all the best William. I would like to see a visionary genius like Elon in medical research to revolutionize knowledge and treatment of diseases.
We lost Rob but got Dave back! :)
I wanna see yall collab
How is your health now Dave? Everything good with the family? Wish you all the best and good health!
Thanks for being back on youtube more regularly,,I enjoy your viewpoint....
TY, hope your wife recovers fine. Löve and hugs from Vienne/Autriche
The future is bright for TSLA!!!
Teslas future is indeed encouraging. However with regards FSD I wouldn't hold your breath.
Tesla is just a car company 😆
@@salty_berserker_channel😂😂 maybe by your traditional standards. This isn’t your dads Ford.
@@juanluismartinez4587 agreed. I was being sarcastic
@@salty_berserker_channel85% of Teslas revenue indeed comes from car sales. Until Tesla commercializes a completed FSD or Optimus robots, or the Dojo computer, or Robotaxi, AI, or licenses their technology to others, yes Tesla is just a car company. Measured in unit sales regulated by transportation agencies, and listed as an automotive manufacturer on Wall St. Shareholders are keen to see a broader car portfolio to drive even more revenue. Yes, at this moment in time Tesla is simply a car company. Teslas performance is judged against other automakers in the segments it serves. It's the profit generated from the car sales allows for funding innovations that augment the transition to sustainable energy but not ready for commercialization. Once the "other innovations" revenue exceeds that of auto sales can Tesla be reconsidered something other than just a car company. Once that happens I think Tesla can then be categorized in a manner that's defined by its revenue.
Hope your wife is doing better Dave. Much love.
I like your analysis which is thoughtful, logical, and concise without throwing huge amount of information. Clearly you have great technical background, and are able to provide concise insight on key technical challenges.
yes people who know a subject domain at a sufficiently deep level don't have to use big words and complex concepts to mask their ignorance
Good seeing more of you Dave. After Rob quit l think you're one of a few Tesla TH-camrs I trust and enjoy watching. Please, stay there, Ok
Cheers!
Thanx for the update Dave
Love you Dave. God bless you and your family. Awesome to see you more often!
Thanks for the update, Dave, take care.
Super exciting for all the work in progress, and the advancements being made. Thanks Dave! Be well, love and light to you, your wife, and family.
Thanks Dave appreciate your videos
Thanks for the time and effort you put into these videos! 👍
Keep up the good work! Love ur content!
So happy to have you back posting regularly after having your wife's health situation more under control. We need your in depth interviews more than ever after TeslaDaily's big loss to the community. Looking forward to updated thoughts with James and Emmet!
At Wealthwise Capital, our goal remains the same. "Mentoring traders to consistent profitability". We have people who started off with nothing but today they're comfortable millionaires. Let that be your story too..
Take some time out your busy schedule to look us up. Remember, it's WEALTHWISE CAPITAL.
Thanks for all your hard work Dave
Glad to see you posting on a more regular basis Dave
Great to see you back Dave!! Hope the family is doing well ❤
Love this streak of posting more often, thanks Dave!!!;)
Thanks Dave and go well with work and family. I love your concise explanations and wise interpretation of Tesla events
Very informative and concise summary thanks Dave! ❤
Thanks Dave; your videos are always top quality.
Spot on. Beautifully and clearly explained! I have experienced the things you talk about in using neural nets for image classification and also with using rule based heuristics, and with setting thresholds and so on. Initial neural training can get you to 90% pd/pfa. Going beyond 95% is hard.
Good to see u more often on TH-cam Dave. Keep it comin
glad to see on YT again.
Thanks for the update Dave! Would be great to see another chat with James about the FSD early results. Would be curious to know at this point what is Tesla's biggest constraint/bottleneck? Is it compute?
Very good delivery Dave. You make good summaries of the important stuff and leave out the whaffle that too many Tesla commentators splutter out time after time. Well done, again!
Thanks Dave🙂
I love you, Dave Lee. Best to you and Yours.🌿🐇
Looking forward to seeing your next.🌿
Terrific overview 👍🏻
Very interesting Dave, Thanks. :)
Good to see you Dave
Hi Dave, It is a big responsibility, but you deserve the added success of likely inheriting a large share of Rob's Tesla Daily subscribers, now that he has "retired" from it to focus on his investment endeavor. Good luck Dave!
At Wealthwise Capital, our goal remains the same. "Mentoring traders to consistent profitability". We have people who started off with nothing but today they're comfortable millionaires. Let that be your story too..
Take some time out your busy schedule to look us up. Remember, it's WEALTHWISE CAPITAL.
@@felixandrew1781boooo… go away Felix
Great.
Thx
☮️🇺🇸
I see an analogy with a driver who has now passed their test but needs experience.
.
The learning experience will now be exponentially fast as more data feeds back and is relayed back to rhe collective.
It's not just a single driver learning.
Great comments Dave. I especially liked your thoughts about breaching some as-yet-to-be-customer-perceived-and-soon-to-be-Tesla-adjusted FSD Beta 12 confidence planning level. This sort of makes me feel like I've been piloting some vague sort of FSD imposter these past years.
Hope you and yours are doing better.❤
I like your take Dave. Thoughtful and insightful.
Do you think it would be helpful for adoption if fsd 12 talked through its thought process and confidence levels in real time to the driver? Like taking your driver's licence you have to give a running commentary, identify hazards and talk through solutions
Thank you
Great vid!
Is DOJO being used for the training?
Thanks for the valuable content! Tesla’s changing the world!
This is a major step
Love your videos:)
Dave, what do you think of the current inference hardware capabilities? Does it have enough “capacity” to drive perfectly and handle all possible corner cases?
Glad to see you back more. Hopefully this also means something worth covering is on the horizon, like ramping up AI. Fsd more positive, and maybe even a buyback. And hopefully ramping the new Model3s
The Neural net outputs massive csv files to the cars on board compute. I wonder if Tesla engineers have an initerpreter for those csv files to be able to see what the neural net is telling the car to do?
Looking forward to trying the awesome FSD V12 on my Model 3!!!
QUESTION:
When you say “confidence” in relation to full neuro net FSD;
Would the computer confidence level be measured by it recognizing that it suddenly has a lack of planning in certain situations or could it be a lack of video data for certain situations, a combination of both or some other issue like lack of computer power in the onboard computer?
I understand that the onboard computer in the vehicle runs comparisons and if there is a conflict it would “give up” as well…Is this correct?
Thanks for the great content Dave.
First----- Dave can you do a deep dive on Palantir's AIP?
Aren't they using the existing rule based system to train / verify the N2N version?
Will autopilot improve with v12?
Sorry if this has been asked here, but will every Tesla eventually get the FSD visualization on the screen? I absolutely love the information it displays as well as the option full screen. But not worth the 12k or 200 a month.
I wonder how long integration and regression tests will take.
I was most impressed in how it handled the emergency vehicles. Navigating around them might have been just normal obstacle avoidance, but given the slower speed I interpret it that it knew to be more cautious, which is a big improvement.
Imagine your car starts to hallucinate... haha
My average fsd trip requires at least 2 interventions and have to contend with sudden stops and misjudging traffic lights…..too many fails. The only thing that really works for me is the basic autopilot on highways. That’s it.
The last update a couple months back had many posting videos of them driving with FSD -- it didn't go well. There is a guy in San Francisco that posted a video, about 15 minutes in length, of his driving around SF using FSD. It should be noted that SF is home turf for Tesla so if there's anyplace FSD should do well it's SF. A few minutes into the video and FSD abruptly turns right heading for a parked car -- the driver had to yank the wheel to avoid hitting the parked car. A short while later FSD aimed for another parked car once again requiring the driver to make a violent jerk of the wheel to avoid hitting it. Not long after that, while descending a hill, something common in SF and many other places, and when the car approached an intersection with a stop sign the car did not stop and rolled right through the intersection without so much as slowing down -- this wasn't a California rolling stop where the driver slows but doesn't stop before making a right turn, the Tesla under FSD drove STRAIGHT THROUGH the intersection without making a right turn or slowing down -- had a child been riding his bike through the cross street at that moment he might well have been run over. In the remaining minutes of the video there were two more instances of the car aiming for parked cars.
So, in this one video there were five cases of highly dangerous behavior with one of them, the stop sign blow through, that could have resulted in death. But yeah, the new FSD is perfect!
I've working in robotics since the mid 80's and I know a thing or two about it -- FSD IS DANGEROUS AND HAS KILLED PEOPLE!
Please tell us where FSD has killed people.
Promises, promises. Dave, why do you think it is, after 7 years since its launch, they still don't have a factory to build the Semi? It could be another 7 years before I see my Cybertruck.
One thing I'm unclear on is the hardware on which most of their data was collected. I assume that it was hardware version 3, since that's what most of the existing fleet is using. On the other hand, I hear occasional reference to the improvements in hardware version 4. Will all the billions of miles collected using hardware version 3 be adequate to train a version that will ultimately be running on hardware version 4? And for those of us who bought FSD years ago be stranded because the hardware isn't adequate?
So is Tesla's v12 FSD better than the already existing Merc's Level 3?
Well, Omar is basically half-Tesla employee
Red hands was heavy rain I believe
what is the difference between Tesla FSD to the ones from Chinese car makers like Huawai and Baidu?
Hi Dave, Is it logical to expect Tesla to be updating the FSD model in the future, but not the underlying code? Or at least we should expect the code base to be infrequently updated. Would that be an indication that Tesla is moving closer to the reality of robo taxis and a fundamental shift in the company?
👍👍👍
filling the gap
Maybe an episode with James Douma is in order?
Why? What make you think he knows whats going on.
I LIKE IT
One of the most interesting parts Elon discussed is that they needed to over-ride the AI and make it actually STOP at stop signs, but 90 percent of humans do not do this, so it wanted to roll through them slowly.
Good Video. However, my prayers go out to your lovely wife😊
ok so thats this year's challenge - and next year? and one after that? ))
All the best wishes Tesla bro
There should be alot of End to End neural net improvements as both FSD and optimus robot developments for real world applications continue their respective development paths. 😂👍
Finally hit it. This FSD thing will be in every car one day
As a Tesla Model Y owner who paid $10,000 for FSD 2 years ago
I look forward to V12
V11 sucks...
*MAYBE* we'll see some positive comment instead of constant trolling?
.
(Not holding my breath on that one)
Technically this is great progress and news. The key question though is how long will it take till FSD will drive a car for a month or so, in busy places, with no interference. The fact that Tesla had to change it’s approach completely confirms my opinion, this change is only the first one of many more changes and many more years.
V12 is upcoming ?
2 weeks
It looks promising. But how about rain, fog, snow - and especially Europe’s narrow, crowded and often illogical roads. My ultimate benchmark is how it’ll cope with a single track road with passing places…. I really want it, but concerned about how long it’ll take.
Were you confident in those situations within 1 year of passing your test? (Honestly)
.
FSD has to *encounter* the condition, probably 100+ times across the available fleet before it *starts* to learn the 50+% appropriate response.
.
Then it's a case of refining the feedback, trending towards 100% correct.
.
If it takes a human with a single input/ feedback (themself) 10 encounters to understand and learn a reasonable response, and the human sees the conditions one time each month, "confidence" takes a year.
.
In the case of something like snow, many drivers NEVER gain sufficient confidence or ability to be safe, depending where they live.
.
FSD has an advantage in that case since there's likely to be snow somewhere on the planet all the time and that input will be relayed and "learned" by all vehicles.
.
The same applies to a "single track road".
.
The questions remaining are the ability to interpret in the first few encounters and rate of correction for any mistakes/ driver corrections.
.
I suspect those will be far superior to the previous method.
The edge cases are going to be very difficult to solve imo. For instance, a driver will know to step on the accelerator and get the hell out of a dangerous situation. Fsd will not move if someone stands in front of the car. It may reverse or slowly try to go around. But it will not recognize the suspect as a threat. Just one of countless scenarios that need to be figured out. How about following police orders during an emergency situation. The cop may ask you to do a u-turn and go the other direction. Or something else unusual.
It is going to be the same old FSD not good enough to remove the wheel or have Robo-Taxi mode.
Ultimately you are stating that change to neural networks still doesn't or cannot give up control, because of the simple reason that the confidence is ALWAYS going to be a variable, and never a constant.
I think you got that back to front?
With all due respect - the end to end neural nets was supposed to be implemented in v69 last year. Every new version, tesla says they removed even more code. How much code is actually left? Noone knows...
How do they incorporate 'stop for a stop sign' w/o coding for it when it's trained on real life data, where nobody completely stops?
I just answered your question on X.
My quick take: I think Tesla is being careful of what it's training v12 on specifically for reasons like this. If you want the car to stop at stop signs then you need to feed it a lot of video showing drivers stopping at stop signs (and exclude videos of drivers not stopping at stop signs). So, there's going to be curation going on in terms of what kind of video data is fed into the neural nets for training. You want to feed it video data that comprehensively represents the type of driving you want out of v12.
In addition to this curated training approach, Tesla might also employ post-training techniques like reinforcement learning. This method involves fine-tuning the FSD model by rewarding desired behaviors (like making a full stop at a stop sign) and penalizing undesirable actions.
This is a huge achievement moving toward the way our brain controls our movements rather than having a rules based interpretation layer between the neural net and the hardware. This also bodes well for their Optimus robots which also use this approach and are improving very rapidly.
Why Tesla is not offering FSD at lower subscription like $99 until FSD become full production version ?
We need more cars on the road.
Please have James Douma back. I want to know how Tesla fixes the disenagements on V12.....if they don't have enough 'GOOD examples. Bottom line how do they fix the critical events.....can't just throw some coding at them like in V11. Please have him back. Thanks.
AI or AGI is a huge topic of concern lately. If FSD is heavily based on AI I'm pretty sure the government and various, safety agencies will have something to say about that. Elon himself strongly recommends government controls on AI.
😍
When FSD really could determine that it can’t handle a situation well, then that’s the first big step to get out of beta. No one expects it to be perfect. We expect it to tell the drive to take over control in time.
This is FSD beta version 0. replacing 300k lines of code means the thing was never gonna work and they're basically staring from scratch.
Would love to see ai's compete in an F1e type situation.
A series based on production cars would be more fitting, like DTM or GT3 or something. Have drivers race for an hour, no battery swaps, only fast charging. Then start a new race with the positions determined by the results of the human drivers. Let AIs drive for another hour and have the points for drivers, AI and total points for the team. Would watch something like that rather than F1
Guards have been removed, what can go wrong
What I understand is Tesla ditched all past years of FSD coding and learning, since the v. 12 works with a totally different end to end AI approach, am I right? Isn't it a huge waste of time and resources? And, are the competence closer, due to v12 was started from scratch and was build in a few months?
Why not build it outside of tala ans own 90% of it rather than 10% of it inside tsla?
What’s wrong with the stock 😢
It's going to be flat (~$200 +/- $100) until FSD is out of beta, or robots are being sold or used internally to cut costs, or car/energy revenue+profits go way up (like >2x).
@@ordinaryhuman5645 I hate that I agree… I also think that this being an election year, and the way they’ve gone after Elon for freeing speech when he bought X, is a contributing factor.
The business is getting better all the time.
It’s just pre-results silliness. No point worrying about the price for at least another 5 years.
Tesla has so many products in development which the Wall St Wankers ignore, that one day they will be forced to recognise them and the price will adjust accordingly. Optimus will probably start to make sense to them soon, following the deal between BMW and Figure Robotics.
you need to be patient. Diamond hands. Benn holding since 2018 and adding. When you are down watch a Ron Barron interview.
@@jeweleratlargeWell, when you accept it for what it is, you can stop being hyperfocused on the trivial short term ups and downs.
L4 or L5 is here! Promised last July by Bossman! Robotaxi’s for all! Appreciating assets! Tesla Network!