ADPVVISOUTH
Member
I agree wholeheartedly. Tesla and SpaceX. These two companies have already crossed two major barriers in two different industries: electric cars and re-usable rockets.Apologies for my earlier response--upon rereading it read a little bit offensive.
Many of these things have been rehashed here and there, but there are a few reasons why GM's claim of having such functionality in 2 years is complete nonsense:
1. Supercruise requires intense 3D mapping of an entire route. Aside from the incredible cost and man-hour effort needed to 3D map and groom such vast expanses of roadway, there are also the issues of construction areas which dramatically change the roadway...anytime construction pops up (for example on an interstate where a lane is temporarily routed to an oncoming traffic lane), these areas need to be updated/maintained. Doing this for just interstates is already a costly and time-intensive task. Doing it for all paved roadways in just the U.S.? Extremely time-intensive, extremely expensive. What happens in the time after a construction project starts and GM comes out to map that section of the roadway? Do you think a construction project is going to wait for GM to come out and map their roadway changes?
2. Such a 3D map data for the US as described in (1) above would take a LOT of data. No way it can be stored onboard the vehicle. So it would have to be continually downloaded to the car as the vehicle drives around. The bandwidth needed to continually download such a high-definition map, and the cost, would be prohibitively expensive. Add this cost on top of an expensive sensor suite and lots of processing hardware. Are you ready to buy a $200,000 Chevy Impala?
3. No matter how many lidars, radars and HD radars you have in your system, you still need a fully-developed neural network for vision. Some examples why:
OK, so a vision-based neural network system is required. How do you develop it? You need several things to develop a good NN system:
- Radar and lidar can't read signs. Yes, you could put all known signs in your mapping data. But signs change all the time. New stop signs get added and removed. Speed limits change. Detour signs get temporarily added for construction zones. New "No Turn on Red" signs get added/removed by municipalities all over the country. All Radar/Lidar sees is a flat metal plane. When a construction company puts up a detour sign, the vehicle MUST be able to read and interpret that sign with vision.
- Lidar and radar can't read stoplights, temporary lit signs (such as temporary speed zones/school zones...speed limit xxx when flashing, etc).
- Lidar and radar have limited ability to interpret hand signals or other visual cues, whereas vision has already demonstrated the ability to do this.
a. Tons, and tons, and tons, and tons of data. How do you gather this data? GM does not have a connected fleet capable of sending images/video back to them for processing and inputting into a neural network. They have Cruise, but Cruise is a relatively small outfit and is only capable of collecting limited data in a very small geographic area.
b. The ability to cultivate the RIGHT kind of data. You can't just throw random data at it and expect it to get better. So you need to find data that addresses scenarios you're trying to solve. How does GM request specific examples of video/imagery from their fleet? They can't.
c. Tons of grooming. The data you feed into the network for training needs to be cultivated and tagged (automatically if possible for much higher productivity, as Tesla is developing with Dojo).
d. Tons of processing power to solve the NN iteratively. Does GM have supercomputers processing their networks, or an auto-labelling system like Tesla has been working on for years? Tesla has among the top 10 or 20 (depending on how you measure it) supercomputer clusters in the world. What have you heard about GM's supercomputer? (That's right, they don't have one).
There are more reasons, but this will give you a good start to be skeptical.
You are right in that Elon was overly-optimistic on the FSD timeline. I said in 2014 on this very forum (my initial post is still searchable) that it would probably take about 10 years to develop level 5 autonomy, which would bring us to 2024. Looks like my estimate is going to be maybe a year or two or three too soon, but in the ballpark.
So while you are correct that Elon thought FSD progress would be way faster than it has been, I think it's also fair to say that Tesla (and perhaps Mobileye--although their data engine is less capable) are the only people working on a true scalable system. I'm still confident Tesla will get there first, but it will take a few more years.
There is no argument. It’s not even close.
