Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Autonomy Investor Day - April 22 at 2pm ET

This site may earn commission on affiliate links.
Agreed that he has at least responded yes on Twitter when asked if 2.0 would have FSD. However, he tweets all kinds of crap.

Yet on the carefully scripted presentation today several slides mentioned HW2.5 —->HW3. If 2.0 was included it would be logical to either state HW2.0/HW2.5 or just HW2.

This may not turn out to be a big deal, but as a 2.0 owner with FSD purchased, I don’t think it is stupid to be a little concerned. But if it does turn out to be a lie I can always join the lawsuit and help some Bay Area law firm get rich from the settlement.

I have 2.0 and purchased FSD just last week. I am not concerned at all!
 
  • Like
Reactions: willow_hiller
Watching the presentation, and hearing testimonies from people who actual rode in a Tesla with the latest FSD, I think it is pretty certain that Tesla has good FSD now, the only question is can they get the reliability up to where it needs to be to qualify as L4/5. But Tesla has self-driving. Their methods are working.
There are probably twenty companies that have good FSD by that metric. It’s all about reliability and edge cases.
 
  • Love
Reactions: jsmay311
18A7869F-7B18-4B3C-8E3C-D096BCD76B02.png
I have 2.0 and purchased FSD just last week. I am not concerned at all!
 
One complaint that I do have about the event is that they did not show any demos during the presentation but left it as something to share later. I think it would have made the presentation better if Tesla had included the demo video as part of the event itself.
 
  • Like
Reactions: Enginerd
No you can't compare a chip created for ADAS versus an entire board with multiple chips. That's completely misleading.
You compare to the Level 4/5 board versus your board.
Its Pegasus versus FSD Board not FSD board versus a single chip.

Plus Xavier is 30 TOPS not 21.

Nvidia says the GPU is 22 TOPS: NVIDIA Jetson AGX Xavier Delivers 32 TeraOps for New Era of AI in Robotics | NVIDIA Developer Blog
The Jetson AGX Xavier integrated Volta GPU, shown in figure 3, provides 512 CUDA cores and 64 Tensor Cores for up to 11 TFLOPS FP16 or 22 TOPS of INT8 compute, with a maximum clock frequency of 1.37GHz.
 
Question is how are those guys planning to get much, much better ? Hand coding or NN. If NN how are they getting data.
Everyone is using NN for their vision systems, what was interesting was that Tesla is exploring using NN for driving policy. Google has images of every single road in the US so I'm betting that their vision system has plenty of data and is working pretty well. Using a NN for driving policy is certainly worth a try and we'll see if it works! I have a hard time wrapping my head around how one would debug such a system.
 
Arrghh.

I thought we'd see the test drives, but they NDA'd those.

Bah.

More than anything else the test drives would have told us how far along they really were.

Edit: They didn't do an NDA, but they didn't allow the riders to record any video. We do have some really descriptive experiences though like this one.

 
Last edited:
  • Like
Reactions: Enginerd
A few things that still are unclear for me:

At some point depending on the neuron counts in your network, adding more data to the network will average out other data and cause edge scenarios to blur our. It becomes underfitted. How confident are they they got the capacity for a large enough network?

From Karpathy's talk they don't just add new connections when new data come in. They also modify existing connections with every new datum comes in. I assume eventually mostly the later part is working and NN size will stop increasing.

Also I didn't like Elon's input of an end-to-end video in - steering wheel/pedals out network. They said themselves a lot of situations are easier to solve with a heuristic approach, while vision and special judgmental cases are best solved with NN. I'd leave it like that.

An approach like that would also imply different datasets for each country/state, or country labels (which implies a much bigger dataset, NN and processing power in car).

Some places you can turn right on red light, some places you can't. In a heuristic approach you just turn of a flag and that if-sentence stops you from breaking the law. Your neural network tells you you're in an intersection staring at a red light waiting to turn. Sew it together and you got a good solution.

Elon said they did not find HD GPS mapping work but it could be used to decide locality and what data set to use. I guess that means different countries/areas will operate on different NN sets. That does not seem to be an issue to me. You only need to store data the car needs.
 
More than anything else the test drives would have told us how far along they really were.
What I got from the presentation is that they're really not very far along but they believe that the infrastructure they've set up will allow them to leapfrog the competition.
I think they'll probably hit a wall jut like everyone else but I'm a pessimist when it come to FSD :p
 
  • Like
Reactions: jsmay311