Here is a good video that shows an example of AEB, which I thought one person said was disabled in the beta, for a bicycle that looked like it might dart in front of it:
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD rewrite will go out on Oct 20 to limited beta
- Thread starter diplomat33
- Start date
Knightshade
Well-Known Member
I have to say, I never, ever, even slightly missed not having a dash cluster in the 3 until seeing the current Beta FSD visualizations full-cluster on an S.
Here's an interesting sequence of screenshots driving up Gough crossing at Washington in San Francisco as not only is it cresting, the lane shifts. Autopilot happens to be following a vehicle at 13mph, so that might also help it change it's predicted path from straight to slightly right, but I wonder if it was clear set at 30mph, would it slow down to react to the lane shift in time?
You live in a binary world and I live in the real world so we are just going to have to agree to disagree here.
I was driving in a French town a couple of years back and came up behind a stretch of road solid with parked cars along its length with no other traffic around. I pulled into the oncoming lane to pass the parked cars. Half way through passing them I saw the traffic lights ahead and suddenly had oncoming traffic heading straight towards me!
And the line of 'parked' cars I was trying to pass started driving forwards. It's not just difficult for FSD to spot the difference between vehicles parked / stationary / queing to turn.
Last edited:
Knightshade
Well-Known Member
No, I live in a world where the current Tesla owners manual says:
Model 3 manual page 100 said:
I'm fully aware of what the manual says. Quoting it doesn't change my mind. Not even a little bit.
".... Do not use Autosteer on city streets, in construction zones, or in areas where bicyclists or pedestrians may be present...."
Otherwise you will get to see the latest FSD / AP features we keep bragging about and continue to develop!
Whatever the book says, Tesla are clearly encouraging existing non-(new) beta use of city based AP features.
Exactly.".... Do not use Autosteer on city streets, in construction zones, or in areas where bicyclists or pedestrians may be present...."
Otherwise you will get to see the latest FSD / AP features we keep bragging about and continue to develop!
Whatever the book says, Tesla are clearly encouraging existing non-(new) beta use of city based AP features.
I mean - Hitachi also claims this thing is a back massager yet we all know what it's actually used for.
willow_hiller
Well-Known Member
Maybe I'm confused about what Dojo actually is, but I find Elon's latest tweets about it odd: https://twitter.com/elonmusk/status/1325445345275416578?s=19
Despite FSD beta relying on 4D training right now, he says Dojo won't contribute until 1 year from now.
Despite FSD beta relying on 4D training right now, he says Dojo won't contribute until 1 year from now.
DanCar
Active Member
Dojo is just a new computing platform to replace the current GPU/TPU clusters. It isn't any new NN/learning technology. (Its only purpose is to speed up the training process.)
Oliver Cameron has another really informative deep dive into Tesla's FSD Beta:
https://twitter.com/olivercameron/status/1325559446433529856
https://twitter.com/olivercameron/status/1325559446433529856
powertoold
Active Member
Isn't the whole point of Dojo to reduce watt per CPU operations so that Tesla can save money in the long run, since they'll be doing lots of video training to improve FSD over time?
There's nothing stopping Tesla from paying third-parties for training until they get Dojo up and running (if ever; it's not clear if Dojo provides a cost benefit as new CPU/TPUs are released by third-parties).
There are significant costs involved with training in datacentres such as Azure. Not on small scale, but Tesla has a huge dataset, big network and rapid/continuous iterations.
Seems Tesla wants a reinforcing feedback loop where the training data/labeling maintains itself based on real world feedback. That's mostly a software problem. A very hard problem since you need an automated scoring/validation mechanism.
I'm not sure how they solved that problem, but I suspect they found out it requires a lot more hardware resources than their current training based on manual labels. Hence the requirement for Dojo to keep costs at a manageable level.
The tweet does somewhat clarify what he was referring to the "rewrite" in that it was primarily the labeling software. This makes the training data quality much improved compared to previous individually labeled photos. Sure, there was probably some changes to the HydraNet to provide additional outputs with the better training data (e.g., birds-eye-view / temporal aspects), but the core structure of the neural network didn't need to be completely rewritten.
If one looks at the progression from AlphaGo (networks trained primarily on human data) to superhuman AlphaGo Zero (networks trained on self play), the network structure is relatively similar (both are "looking" at a Go board and predicts how good a move or position is) but the performance is much improved in the latter because the quality of training data.
Last edited:
Similar threads
- Question
