-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
OFFICIAL BUTTON WATCH
- Thread starter Drrobmiller
- Start date
powertoold
Active Member
I wouldn't waste my time arguing with someone who doesn't admit they're wrong.
One item I find interesting is that FSD beta drivers in the San Francisco area seem to have far few disengagements then drivers in other parts of the country. Elon seemed to confirm that with his tweet regarding "overfitting" the area. (whatever that means). So there must be something to this thing called "data".
We'll know alot more when V10 is released and whether it represents a significant improvement or not. Until then I'm just waiting. Arguing is pointless until after the V10 videos are out IMO.
We'll know alot more when V10 is released and whether it represents a significant improvement or not. Until then I'm just waiting. Arguing is pointless until after the V10 videos are out IMO.
Last edited:
powertoold
Active Member
I don't think that's true. Whole Mars "games" fsd beta (uses accelerator a lot) and also lets it do sorta risky maneuvers. And he's mostly the one posting all the SF videos. There's a recent SF video from AI Addict that is pretty bad.
caligula666
Member
Except it is not the case: being available in the US (which it is far from being...), doesn't make it available in other geos for obvious different rules, regulatory and cultural reasons. But, hey, your angry bad boy self-centered style is still funny.
Iain
Member
It was addressed on AI day.. they made it very clear that once FSD is solved in the US it would be easily transferable to any other country.
I wonder how many folks with elderly parents or simply elderly themselves will find FSD V10+ worthwhile even if it's not level 5 to help them get to doctor appointments or just in general allow them to continue driving longer? Especially in suburbia. I for one remember when I had to literally take the keys from my elderly mom (88) and as you can imagine she wasn't too happy at the time. In hindsight a good solid FSD version would have let her keep the keys a bit longer. Stuck in the house is not a good thing for someone who is used to getting up and about.
Who cares from the scope we are asking about? FSD beta can “operate” anywhere in the United States except NYC.
I guess you probably should have pointed out It isn’t everywhere because it isn’t available on Mars yet. Or in alpha centari. Neither is waymo. But who cares. Well, you do because you brought it up.
All those locations, including outside of the United States are outside the scope of any full self driving program being developed in the United States. It’s irrelevant.
Last edited:
I think people underestimate the value in this application. Its similar to NoA. For me, it makes a long trip so much less tiring and then I can focus on watching for animals, pot holes, etc. If it is as reliable as NoA, city streets would be huge for the elderly.
Of course the issue then becomes what is the ratio of people that mis-use it that then cause accidents vs people that it enables them to drive again and makes all of our roads safer.
Unless the L2 or L3 FSD can fail with plenty of warnings and automatically place the car away from harm, it may become dangerous asking people who have already struggle with driving to monitor and react rapidly to a L2/L3 driving assistance system. I would think you want a L4+ system for those situation, so you don't need to worry about any scenarios where quick reactions will be needed by the driver.
That's interesting. I'm not surprised they throw out most of their data. Also they have a limited amount of bandwidth to label it, even with automation. It's still important to collect lots of video of the same scenarios, not just for training, but also validation... but yeah, at some point you're going to collect enough.
Also, as demonstrated at AI day, they are able to query their raw data for very specific scnarios.
And yes, some scenarios require simulation because they either never happen or don't happen often enough. However, even when filling the gaps with simulation, Tesla has a data advantage as they use existing footage to help generate simulated footage, in ways that prevent their NNs from being "overfit" to simulations.
Solid compared to Waymo's operation in Arizona? Depends on the scnario from what I can tell. They're different products and hard to compare. Solid compared to MobileEye's handpicked, vision-only, promotional material? Yeah, Tesla's FSD needs some work.
But solid compared to what people thought was possible years ago? Solid compared to those who still claim FSD is vaporware and doesn't exist, or barely functions and will ultimately fail? Yes, compared to that I think it's pretty solid.
I'm personally impressed, as a software engineer, watching what they've built function at all. It's exciting and fascinating to me, and I'm actually inspired to the point where I'm working on transitioning into that industry. I've ignored autonomous cars for a long time because it seemed like an insurmountable challenge - but watching FSD beta, out in the wild in novel situations, actually make a vaguely reasonable attempt at driving... and on consumer level hardware, too... I think it's pretty neat. I suppose that's just my opinion though.
I feel like I have a good handle on telling the difference between incomplete software that has the important bits in place, and incomplete software that is destined to be abandoned, and I believe FSD is in the former category.
I guess this isn't meaningful to debate without specific quotes. At least in this forum, I have certainly felt a shift in the last few months from people debating if vison only was possible at all, to discussing specific camera placement and specifications. To me that shows that people went from skeptical of vision-only, to seeing how it could work.
Of course! MobileEye's NY promotional demo was fantastic. I also liked watching Zoox's SF demo, as well as customer footage from Waymo's rides (though of course thoss latter two use lidar and HD maps).
This topic is about Tesla though. I think Tesla's situation is uniquely fascinating to me because I'm watching these FSD videos where they all use software running on the exact hardware I have outside sitting in my driveway; in a thing that I've already purchased and already enjoy using every day. I'm very anxiously waiting to see what Tesla will finally be able to deliver on that. So FSD tends to receive more of my attention.
I'll admit my bias towards Tesla. A success for Tesla means that I gain access to a product I'm fascinated by. But I've been very clear in the past that I want to see them all succeed. I personally think other companies will achieve profitable robo-taxies years before Tesla will (assuming Tesla robo-taxies are even ever a thing). But Tesla (IMO) will probably dominate ADAS for years to come.
Elon Musk specifically mentioned during AI Day Q&A in response to a question to "improve one component along the AI stack… biggest impact for FSD performance:"
"In the short term, we need all of the nets to be surround video. We still have some legacy…"
This most likely involves collecting new data and getting it labeled (automatically) as the existing networks were trained on 2d single camera images. Tesla is able to move fast and make these constant iterations /because/ they have the fleet data collection capability. In some sense, they're able to throw out the old approach and not be slowed down by sunk costs of the existing implementation and trying to squeeze out value of previously collected data that might be lacking video from other cameras necessary for surround video training.
I'm not sure I follow. Moving from single frame training to training on video shouldn't require different data? Just instead of sending single frames for training, they send in the entire video clips. I guess I'm assuming that they've been collecting video clip data from the fleet this whole time, and not just individual frames.
I think he was responding to someone who asked. Not you.
What your saying is you don't like silly timelines? Then don't look at them.
FloridaJohn
Member
I think he was pretty clear. “Solid” as measured by number of disengagement’s per mile.
I remind you that the introduction to FSD was a Tesla promo video in 2016, showing the car pulling into a parking lot, reading a no parking sign, and picking a new spot, and parking with nobody in the car? You know, the video that is still on the Tesla site on the AP page?
Or Elon saying they would do a cross country drive with no disengagements by the end of 2017.
It's not very solid compared to what Tesla thought (and advertised) was possible in 2016. What was sold was door to door, no disengagements, no action needed from the driver, and yes, that doesn't exist today even with "FSD" beta, which is actually city streets autosteer.
I have the benefit of only starting to pay attention recently. I admit that the experience of those who have been watching since day one is filled with a lot more disappointment. I'm not sure what I can say to bridge those two experiences, other than to simply point out that they are different.
As someone that has been around for a while, I hope your experience is different. However, just remember that Tesla has been making self driving promo videos since 2016, and many people bought back then because it seemed like it was actually possible in 2017 too. 4 years later just because you see a Youtuber make one today doesn't mean they are all that close to allowing the car in your driveway to actually run this code, despite Elon saying it's "soon." Using any kind of history from Telsa, the only way to accurately know when something will ship is when it's actually on your car. Everything before that is hype.
I have a 2019 car- and guess what, this code won't run on it, despite Tesla telling everyone all cars had the HW needed. We already know Tesla is working on HW4 and hyping that. There's as good a chance that this will never be released to HW3 cars as there is that it will.
And we only get to see the videos the testers chooses to publish. How many failures on video is discarded and never shown?
While this is true, most of the beta testers seems to try to put the car into situations to test it on the videos. I find it just as likely that there are a lot of zero intervention drives that we don't see because they're just that boring.
Similar threads
