Read on Twitter
1/ Another FSD thread.
@karpathy points out that $TSLA's FSD competitors use high-definition maps to plot out all drivable space
These cars don't need to see Stop Signs because they've already programmed/baked in all the stop signs
Let's discuss, $TSLAQ https://twitter.com/ElectrekCo/status/1273764197126504449
@karpathy points out that $TSLA's FSD competitors use high-definition maps to plot out all drivable space
These cars don't need to see Stop Signs because they've already programmed/baked in all the stop signs
Let's discuss, $TSLAQ https://twitter.com/ElectrekCo/status/1273764197126504449
2/ A robot car doesn't need to "see" stop signs if it knows ahead of time where all the stop signs are.
Similar, but the car doesn't need to decipher lanes or any signs. It's been pre-programmed that "this lane is the left-turn only lane." Or "this is where I have right of way"
Similar, but the car doesn't need to decipher lanes or any signs. It's been pre-programmed that "this lane is the left-turn only lane." Or "this is where I have right of way"
3/ And @karpathy is right that this isn't ideal
No one can map out all drivable space and keep it 100% up-to-date all the time.
It would be ideal for cars to be able to see the signs and infer their meaning.
Stop sign --> stop here
Left turn only --> leave lane to go straight
No one can map out all drivable space and keep it 100% up-to-date all the time.
It would be ideal for cars to be able to see the signs and infer their meaning.
Stop sign --> stop here
Left turn only --> leave lane to go straight
4/ So why do these companies do it?
Because it's the critical path to getting to market.
Driving on ordinary streets is already a monumental challenge. Without the map, you're introducing another layer of complexity: inferring the world around you.
Because it's the critical path to getting to market.
Driving on ordinary streets is already a monumental challenge. Without the map, you're introducing another layer of complexity: inferring the world around you.
5/ So these companies have all arrived at a similar conclusion:
Step #1: Get the car to drive perfectly with a 100% perfect map
Step #2: Get the car to drive perfectly with no map
The progression here is obvious.
Step #1: Get the car to drive perfectly with a 100% perfect map
Step #2: Get the car to drive perfectly with no map
The progression here is obvious.
6/ The second benefit of this approach is that it enables FSD companies to get to market much faster. You can launch as soon as you've solved step #1.
There's billions up for grabs in just a few major US cities. Establishing your brand in major cities helps the broader rollout.
There's billions up for grabs in just a few major US cities. Establishing your brand in major cities helps the broader rollout.
7/ So while @karpathy is correct that pre-mapping in FSD isn't ideal...
... he's very wrong to imply that it's a poor decision.
Mapping is a relatively cheap solution that will enable these FSD companies to start seizing transportation market share in major cities. Lots of $$$
... he's very wrong to imply that it's a poor decision.
Mapping is a relatively cheap solution that will enable these FSD companies to start seizing transportation market share in major cities. Lots of $$$
8/ And it may even be technically a better approach.
With the map, the FSD companies can stay very focused on discreet ML problems.
Once they've solved a very safe FSD car, they can then try to go to the next level and drop the map.
With the map, the FSD companies can stay very focused on discreet ML problems.
Once they've solved a very safe FSD car, they can then try to go to the next level and drop the map.
9/ Similar point can be made with LIDAR:
If LiDAR is not necessary, then these companies will eventually be able to drop it. No reason they can't just say "ok, we're training just off the cameras now."
If LiDAR is not necessary, then these companies will eventually be able to drop it. No reason they can't just say "ok, we're training just off the cameras now."
10/
Point is this:
Once you have solved a safe FSD vehicle that can pick up and drop off passengers...
Once you have SOLVED the world's greatest AI challenge...
Your company moves from "solving" to "optimizing." All other AI challenges get vastly easier.
Point is this:
Once you have solved a safe FSD vehicle that can pick up and drop off passengers...
Once you have SOLVED the world's greatest AI challenge...
Your company moves from "solving" to "optimizing." All other AI challenges get vastly easier.
11/
But $TSLA will never solve this, because the FSD challenge is one of the hardest challenges of our time, with some of our brightest minds trying to solve it.
and $TSLA can't retain talent
$TSLAQ
But $TSLA will never solve this, because the FSD challenge is one of the hardest challenges of our time, with some of our brightest minds trying to solve it.
and $TSLA can't retain talent
$TSLAQ
12/ $TSLA's brightest minds realize Elon's constrains create an unsolvable problem. And they leave.
FSD isn't a one-man show. No one person put a man on the Moon.
These achievements come from organizations committed to unconstrained innovation. Elon is antithetical to that.
FSD isn't a one-man show. No one person put a man on the Moon.
These achievements come from organizations committed to unconstrained innovation. Elon is antithetical to that.
13/ To say that Elon Musk and @karpathy, in all their genius glory, are going to solve FSD cheapens the work of the hundreds of researchers that are actually solving the problem.
And it begs the question...
Why aren't their test $TSLA cars on the road testing FSD today?
And it begs the question...
Why aren't their test $TSLA cars on the road testing FSD today?
