So you watched the entire thing before it was taken down? How about throw us a bone and do a TLDR?Great thread Karpathy talk Tesla fsd. A must watch the YouTube was taken down temporarily, I hope.
You can install our site as a web app on your iOS device by utilizing the Add to Home Screen feature in Safari. Please see this thread for more details on this.
Note: This feature may not be available in some browsers.
So you watched the entire thing before it was taken down? How about throw us a bone and do a TLDR?Great thread Karpathy talk Tesla fsd. A must watch the YouTube was taken down temporarily, I hope.
Several years back, @TrendTrader007 used to post here regularly. Typically would be a message like "Based on the chart, $TSLA poised for huge break out to $XXX" - this is OK, but he'd post these every other week, of course it never happened and he ended up getting laughed at so much that he left, returned, left, returned, then finally left, until he returns again...Is this an inside joke based on his 6 month old videos?
The Twitter user was able to upload clips from his browser cache.So you watched the entire thing before it was taken down? How about throw us a bone and do a TLDR?
Great thread Karpathy talk Tesla fsd. A must watch the YouTube was taken down temporarily, I hope.
CVPR, great, no confidential information then, but would be too technical for Wall Street to understand.Andrej Karpathy Tesla Autonomous Driving Talk CVPR June 20 2021
Direct YouTube link to entire talk:
Ok, thought that were only rumours (or educated guesses - Elon said they would do structural battery packs but as in the first iteration of the planning documents there was no cell production, Kato would have been the only possible source). Judging from the likes your post got that is not the case . Does someone have a source?The plan has been to start with imported batteries from Kato since forever. If they can stick to this looks like they'll be producing cars at around the same time as Austin. Not bad even with three months headstart considering all the permit delays and Austin using five times the personnel for construction
As expected, not there for me (Europe) ...Don’t think I’ve seen this before? Tesla app under Upgrades>Manage Upgrades I’m seeing this:
View attachment 675648
(I already have AP and FSD).
Sorry if this isn’t new.
edit: typo
Direct YouTube link to entire talk:
Edit: This is a must watch. If you only have limited time skip to 3:30, but the rest is very information rich.
Agreed, must watch, not very technical actually, so don’t be intimidated just because it’s for CVPR conference.Direct YouTube link to entire talk:
Edit: This is a must watch. If you only have limited time skip to 3:30, but the rest is very information rich.
The article from @avoigt suggested via a leak that there would be two production lines at Giga Berlin. However I also recall this Business Insider article from earlier in the month that claims that Elon scrapped the second production line during his most recent Berlin visit. Tesla: Elon Musk hat bei seinem jüngsten Besuch in Grünheide Baupläne über den Haufen geworfenOk, thought that were only rumours (or educated guesses - Elon said they would do structural battery packs but as in the first iteration of the planning documents there was no cell production, Kato would have been the only possible source). Judging from the likes your post got that is not the case . Does someone have a source?
Also, if they do both 2170 and 4680 as @avoigt implied, I would expect they would start with 2170 and not jump to structural battery back in the beginning.
Tesla related: UK second hand prices for Model 3 - less than 2 years old (introduced 2019) are down about 8% (Autotrader UK asking prices plus other sources*). Only exotics fare better (and you might have to be VERY special for the OEM to even consider you as a customer). Parkers is saying lower - the lowest private sale being around 78%, highest dealer price 91% after 2 years Free valuation | ParkersA friend of mine might be getting a tesla. I was looking on their CPO site and was amazed at the prices of the CPO cars. They are as much as a new one. There also were not many cars available either.
Agreed, must watch, not very technical actually, so don’t be intimidated just because it’s for CVPR conference.
My notes:
TL;DL:
- Briefly touching vision vs Lidar and mentioned Lidar is not a generalizable solution.
- Now vision is working well that they can start to remove other sensors.
- Mainly focused on the feature of using pure vision to remove radar in AEB(auto emergency braking) to illustrate how Tesla AI works
- Start with data accusation, they have 221 triggers running that will record 10sec video and send back for events like “driver brakes hard on freeway” or “detected lead car slow down but driver didn’t brake” etc.
- Offline auto labeling to get accurate depth/velocity, with benefits of hindsight and more computing, auto labeling could actually get very good quality labels.
- Train net and deploy to fleet in shadow mode.
- Then collect more disagreements and retrain.
- 7 rounds of shadow mode deployment before final release.
- In total 1M clips collected, 6billion labels, 1.5Pb dataset size(some tweets interpreted this as the total size of FSD dataset, my interpretation is 1.5Pb is for AEB feature alone)
- Compared with legacy solution with radars, the new pure vision system avoids phantom brake, detects stopped truck earlier and works more reliably in case of hard braking of lead car(in that case radars lose tracking of the lead car and re-acquired it a few times in one second, produced a lot of noise)
- QA process includes 10k simulation scenarios, 10years of QA drive and 1000years of shadow mode drive. (Next time Waymo brags about simulation we can say Tesla do that too, on top of millions of real world video clips, hand picked from billions of miles of real world drives)
- Tesla training cluster has 5760x NV A100 GPUs, the 5th largest supercomputer in the world. And this is only 1 of the 3 clusters they are building. Next step is Dojo, but not ready to talk about that yet.
- Was asked about whether adding other sensors would be helpful, for example infra cam, answered there is infinite number of sensors you can chose from, but they still think cameras in visible spectrum are the best choice, it has all the information needed for driving, basically it’s necessary and sufficient for FSD.
Demonstrated a lot of speculated Tesla AI moats, data collection triggers, auto labeling, shadow mode, training clusters, etc.
Also shows great progress on pure vision using the example of vision AEB.
Bullish AF!
Release
15M miles
1.7M on Autopilot (0 crashes)
Or only HW3 cars.That must be a very recent release to have so few Autopilot miles, but there does not seem to be a recent release according to TeslaFI (last was 2021.4.18.3 on 10 June).
I watched his talk. It was a lot like I was speculating half a year ago.Agreed, must watch, not very technical actually, so don’t be intimidated just because it’s for CVPR conference.
My notes:
TL;DL:
- Briefly touching vision vs Lidar and mentioned Lidar is not a generalizable solution.
- Now vision is working well that they can start to remove other sensors.
- Mainly focused on the feature of using pure vision to remove radar in AEB(auto emergency braking) to illustrate how Tesla AI works
- Start with data accusation, they have 221 triggers running that will record 10sec video and send back for events like “driver brakes hard on freeway” or “detected lead car slow down but driver didn’t brake” etc.
- Offline auto labeling to get accurate depth/velocity, with benefits of hindsight and more computing, auto labeling could actually get very good quality labels.
- Train net and deploy to fleet in shadow mode.
- Then collect more disagreements and retrain.
- 7 rounds of shadow mode deployment before final release.
- In total 1M clips collected, 6billion labels, 1.5Pb dataset size(some tweets interpreted this as the total size of FSD dataset, my interpretation is 1.5Pb is for AEB feature alone)
- Compared with legacy solution with radars, the new pure vision system avoids phantom brake, detects stopped truck earlier and works more reliably in case of hard braking of lead car(in that case radars lose tracking of the lead car and re-acquired it a few times in one second, produced a lot of noise)
- QA process includes 10k simulation scenarios, 10years of QA drive and 1000years of shadow mode drive. (Next time Waymo brags about simulation we can say Tesla do that too, on top of millions of real world video clips, hand picked from billions of miles of real world drives)
- Tesla training cluster has 5760x NV A100 GPUs, the 5th largest supercomputer in the world. And this is only 1 of the 3 clusters they are building. Next step is Dojo, but not ready to talk about that yet.
- Was asked about whether adding other sensors would be helpful, for example infra cam, answered there is infinite number of sensors you can chose from, but they still think cameras in visible spectrum are the best choice, it has all the information needed for driving, basically it’s necessary and sufficient for FSD.
Demonstrated a lot of speculated Tesla AI moats, data collection triggers, auto labeling, shadow mode, training clusters, etc.
Also shows great progress on pure vision using the example of vision AEB.
Bullish AF!
This is another brilliant presentation by Andrej.Agreed, must watch, not very technical actually, so don’t be intimidated just because it’s for CVPR conference.
My notes:
TL;DL:
- Briefly touching vision vs Lidar and mentioned Lidar is not a generalizable solution.
- Now vision is working well that they can start to remove other sensors.
- Mainly focused on the feature of using pure vision to remove radar in AEB(auto emergency braking) to illustrate how Tesla AI works
- Start with data accusation, they have 221 triggers running that will record 10sec video and send back for events like “driver brakes hard on freeway” or “detected lead car slow down but driver didn’t brake” etc.
- Offline auto labeling to get accurate depth/velocity, with benefits of hindsight and more computing, auto labeling could actually get very good quality labels.
- Train net and deploy to fleet in shadow mode.
- Then collect more disagreements and retrain.
- 7 rounds of shadow mode deployment before final release.
- In total 1M clips collected, 6billion labels, 1.5Pb dataset size(some tweets interpreted this as the total size of FSD dataset, my interpretation is 1.5Pb is for AEB feature alone)
- Compared with legacy solution with radars, the new pure vision system avoids phantom brake, detects stopped truck earlier and works more reliably in case of hard braking of lead car(in that case radars lose tracking of the lead car and re-acquired it a few times in one second, produced a lot of noise)
- QA process includes 10k simulation scenarios, 10years of QA drive and 1000years of shadow mode drive. (Next time Waymo brags about simulation we can say Tesla do that too, on top of millions of real world video clips, hand picked from billions of miles of real world drives)
- Tesla training cluster has 5760x NV A100 GPUs, the 5th largest supercomputer in the world. And this is only 1 of the 3 clusters they are building. Next step is Dojo, but not ready to talk about that yet.
- Was asked about whether adding other sensors would be helpful, for example infra cam, answered there is infinite number of sensors you can chose from, but they still think cameras in visible spectrum are the best choice, it has all the information needed for driving, basically it’s necessary and sufficient for FSD.
Demonstrated a lot of speculated Tesla AI moats, data collection triggers, auto labeling, shadow mode, training clusters, etc.
Also shows great progress on pure vision using the example of vision AEB.
Bullish AF!
Several years back, @TrendTrader007 used to post here regularly. Typically would be a message like "Based on the chart, $TSLA poised for huge break out to $XXX" - this is OK, but he'd post these every other week, of course it never happened and he ended up getting laughed at so much that he left, returned, left, returned, then finally left, until he returns again...
He couldn't handle people criticising his perpetual proclamations and started to be abusive, with stuff like "screw you amateurs, I've more money than all of you combined, go f yourselves", or words to that effect
We never knew how much $TSLA he had, but he implied it was >100k shares pre-split. He was forever flipping in and out of stock and calls. In theory he should be a billionaire by now, who knows...
I follow him on Twitter, he's somewhat schizophrenic, one day he's loading on $TSLA, the next he claims the apocalypse is coming and he has sold everything and moved into cash
Direct YouTube link to entire talk:
Edit: This is a must watch. If you only have limited time skip to 3:30, but the rest is very information rich.