JupiterMan
Active Member
You may not agree with his assessment, but as a leading academic in AI, Kaparthy is pretty damn far from an idiot.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
[uk] UltraSonic Sensors removal/TV replacement performance
- Thread starter randompixel
- Start date
You may not agree with his assessment, but as a leading academic in AI, Kaparthy is pretty damn far from an idiot.
I quote BMW.
"Soon to be released, the very first BMW i7 pushes the envelope further in the luxury electric saloon market"
Unless you know different!
Kaparthy had to go, anyone who takes 5 years trying to stitch a radar feed and 8 camera images together before deciding it's difficult clearly has no practical sense of timelines, or a grasp of reality. The decision to drop radar, if it was as bad as they say, should have been acted upon years ago. If they can't make a decision on such a fundamental quicker than that, what hope have we for anything else? To be honest, any sane company would have checked before shipping cars with the sensors, but thats another story.
JupiterMan
Active Member
He did not have to go.
Kaparthy left of his own accord to go back to teaching, go back to programming and away from management side of things. Musk did not want him to leave and is on the record as saying he is always welcome back. Kaparthy has also hinted he would be open to going back at some point.
Tesla had to invent new NN technologies to solve the self driving problem, they had many glass ceilings as they progressed which meant they had to change tack. Remember no one else was (or still is) doing this with Vision only, so Tesla were inventing solutions, testing and iterating continually.
They dropped the radar once Vision had progressed to a level that they could compare the radar and Vision systems side by side and decided to ditch the radar.
Similar for USS I would assume, now that the occupancy networks are live, they seem to have made the decision on comparing outputs that the USS are now superfluous.
Such decisions can only be made when the new technology reaches or exceeds that of the old, it cannot be made years before.
It is likely that the decision to drop USS was made a long time ago. Kaparthy talked about this decision, so it seems it was made when he was there at Tesla and before his sabbatical, so quite some time ago.
Once made a decision would have been made on a future date to delete the sensors, they would then have to stop USS orders (in the magnitude of about a million sensors a month) and to redesign components and manufacturing processes for the USS deletion launch date.
A lot of factors would have gone in to this, and it seems the software side wasn't quite ready for the USS deletion date, so yes a miss from Tesla's side, but once a decision like this is made you cant simply revert back.
However, It seems FSD v11 is now shipping as from today, so lets see if this build includes the USS replacement code for the the new cars which don't have the sensors. If it does then the miss has only been a few weeks, with very few customers impacted.
Those that believe, will believe
Those that doubt, will doubt
Those that look at the evidence, will know.
I think George was referring to the radar and whether they should have made sure that work before installing it in over a million cars. You have to admit it's embarassing that it took so long before they decided to drop the radar while all that time punting out messages such as "feature complete this year", "Robotaxi coming this year" etc etc. year after year signalling they were on the cusp.
You simply can't have it both ways. How can it be an amazing feat to get to where they are now with so many new technologies invented, so many amazingly cool things developed which take time AND also be sending out the message it was all about to come good. Do you really think that they've spent the last 6 years thinking "just one more thing to solve and we'll be done" only to find when they've solved it, there was yet another thing they didn't know? And if you do believe that is what happened, then surely you have to accept there is a very strong case that after the next "only one more thing to be solved" is solved, there will be yet another unexpected thing around the corner and the current claims of "we're nearly there" are as unreliable as the previous claims.
It also seems the FSD Beta v10.69.3 has been withdrawn as well, as after a week or so it's on no more than about 8 cars if you look across the various firmware tracking websites. Yet another failed release.
Those that doubt, will doubt
Those that look at the evidence, will know.
I think George was referring to the radar and whether they should have made sure that work before installing it in over a million cars. You have to admit it's embarassing that it took so long before they decided to drop the radar while all that time punting out messages such as "feature complete this year", "Robotaxi coming this year" etc etc. year after year signalling they were on the cusp.
You simply can't have it both ways. How can it be an amazing feat to get to where they are now with so many new technologies invented, so many amazingly cool things developed which take time AND also be sending out the message it was all about to come good. Do you really think that they've spent the last 6 years thinking "just one more thing to solve and we'll be done" only to find when they've solved it, there was yet another thing they didn't know? And if you do believe that is what happened, then surely you have to accept there is a very strong case that after the next "only one more thing to be solved" is solved, there will be yet another unexpected thing around the corner and the current claims of "we're nearly there" are as unreliable as the previous claims.
It also seems the FSD Beta v10.69.3 has been withdrawn as well, as after a week or so it's on no more than about 8 cars if you look across the various firmware tracking websites. Yet another failed release.
Let’s be honest, the Vision alternatives while they ‘work’, don’t work as well as dedicated hardware. That to me is not even an issue, because it’s still impressive how far they progressed with Vision and I think eventually Tesla will achieve their objectives.
What really grates my cheese are to me the blatant lies about how far ahead they are with their tech. Vision and therefore FSD, just isn’t ready. That fact is fine, the lying isn’t.
Innovation thrives through ambition. But ambition != over promising.
What really grates my cheese are to me the blatant lies about how far ahead they are with their tech. Vision and therefore FSD, just isn’t ready. That fact is fine, the lying isn’t.
Innovation thrives through ambition. But ambition != over promising.
This is a thoughtful and considered post and I'm doing it a disservice by only quoting this line, but there surely comes a time when, if no one else has adopted the same solution you have, you have to ask if it was the right decision; instead of pigheadedly continuing to (say) use a fork to eat soup.
Zilla91
Active Member
Not wanting to be too disparaging but what has Karpathy actually done? I’m not talking about research papers or anything like that, actual things that are in use today. As far as I can see he’s an academic with lots of theories but the only thing he’s actually done is at Tesla and it’s less than stellar, so much so he jumped ship to go back to opining as an academic again.
SalisburySam
Active Member
…or using a camera instead of a dedicated sensor to determine if it is raining and how hard.
The daft thing is that a camera would be the better way..so long as it was in the cabin, able to see the whole windscreen rather than a few square cm easily marred by dead insects etc that get made worse by dry wipes.
To be fair, Mobileye are doing a comprehensive vision only version, but it looks to be only ADAS, albeit quite advanced (which is pretty much where Tesla are currently trying to get to). 8M cameras (were Tesla looking to maybe upgrade to 4M?) Geeley are using this system now.
Mobileye SuperVision™ | The Bridge from ADAS to Consumer AVs
Mobileye SuperVision™ is our most advanced driver-assist system, providing eyes-on/hands-off automated operation of standard driving functions.
Their Level 3+ systems have a slightly more comprehensive hardware stack including 13 cameras, 9 lidars, and 6 radars. I think its fair to say, thats a signifcantly more comprehensive sensor suite.
Its kind of interesting comparing how Mobileye present their information on the website and the way Tesla do. I know this type of presentation style doesn't necessarily mean the underlying tech is better or worse, but you do get the feeling they've thought things through, talk of mixing perception with driving policy, sharing information, etc. They're doing what Tesla are doing by reading the roads and building a comprehensive map, but they're also sharing that between cars so you know whats arund the corner before you can even see it.
Being good at AI does not mean he is good at understanding of driver needs or car.
AI is one thing, which is fine. Not fine is that he stubbornly pushes teslavision instead of PROVEN RELIABLE and WORKING solutions. And TV is not working.
It SHOULD be easy to make it work, but...
Zilla91
Active Member
Tesla vision continues to cover itself in glory
Wow that's embarrassing.. mirrors what my experience has been though, it's esentially useless. It rarely offers a parking space, it's tuned to look for massive US spaces not our small ones.
Yeah, that's broadly been my experience of auto park too. I'd like to say I'm shocked that it's so bad, but I'm not. It's just another thing that Tesla don't seem to be remotely interested in getting right. The fact it can't even recognise perpendicular spaces is just embarassing.
Quite consistent with my experience sadly.
But kinda jealous of the reverse summon available on the Ioniq. Would love to have that 'remote autopark' option from the app... if the car wasn't doing such an embarrassing job in the first place..
Hadn't seen that one before with attempts at using vision, but saw the previous video versus i4 etc. Seems like vision has made not a jot of difference. EAP / FSD unusable, whereas Hyundai / Kia seems perfectly fine. And can be used remotely.
I think I know where most sensible buyers will look next.
Believe the issue is that the computer took too long to determine what to do when it received conflicting data from Vision, Sensors and Radar. Having only one pure data stream (Vision) allows the Autopilot to react quicker.
Kind of like when you are casually driving, and your eyes/brain tells you you have a clear path, but your passenger is constantly shouting...WATCH OUT!...BE CAREFUL!..
Kind of like when you are casually driving, and your eyes/brain tells you you have a clear path, but your passenger is constantly shouting...WATCH OUT!...BE CAREFUL!..
Similar threads
