Tesla: “Full Self-Driving beta” Isn’t Designed for Full Self-Driving
SAE level 2 is industry jargon for a driver-assistance systems that perform functions like lane-keeping and adaptive cruise control. By definition, level 2 systems require continual human oversight. Fully driverless systems—like the taxi service Waymo is operating in the Phoenix area—are considered level 4 systems.
... If Tesla is really going to achieve fully driverless operation in 2021, that doesn't leave much time to develop, test, and validate complex, safety-critical software. So it would be natural for customers to assume that the software Tesla named "Full Self Driving beta" is, in fact, a beta version of Tesla's long-awaited fully self-driving software. But in its communications with California officials, Tesla makes it clear that's not true.
... Perhaps Tesla will move faster than Waymo and it won't take another five years to achieve fully driverless operation. But customers considering whether to pay $10,000 for Tesla's full self-driving software package should certainly take Musk's optimistic timeline with a pinch of salt.
OK so let's take apart the marketing BS and strip this down.
Tesla has a license for Level 2. End of story. So Tesla can't offer anything other than Level 2. That means even if the software is level 3/4/5 capable, the driver must still be aware, watching the road and able to take over.
Tesla denotes this by the same system it has with AP where you simply have to be in touch with the wheel or touch one of the buttons.
That has absolutely Not One Single Thing to do with the capability of the system, the advanced functionality, whether or not it actually can drive the vehicle or not, what the training regime is or how capable the hardware/software/training environments are with the ability to grow into a Level 4 or Level 5 system. Nor does it denote how long it will take to get there.
I will reiterate a post I made about this quite a while ago which goes into machine learning as opposed to human learning.
Google decided to create a video game playing AI. It took them 3 years to build the AI to the point of "feature complete". During that 3 years Google had been recording, categorising and tagging video game play from some of the best players on the net.
Once the AI was feature complete, google took the 3 years worth of categorised and tagged data and fed it into the AI. It then let the AI loose with the initial data and left it to learn at machine speed form the initial start point.
3 DAYS later Google had a viable game playing AI.
What is Tesla doing? Tesla is using the onboard hardware to drive the vehicle. In order to do so, it has to read the available video and radar data in real time and project a "drive path" through the roads, other road users, traffic, obstructions and 1001 other edge cases. But the Tesla computer is far more powerful than the effort needed for that so the computer is constantly presenting alternate paths and alternate interpretations of the objects it sees around it.
Meanwhile, back at the ranch as they say, Tesla is building a data monster called DoJo. It is an AI training computer which is designed specifically to be able to accept all the data from these FSD vehicles and even all the other Tesla vehicles where it is running in shadow mode and comparing it's performance with the actual decisions of the human drivers.
Where DoJo can auto label, it is madly crunching its way through the humungous amount of data being produced. Where there is doubt, DoJo will throw out images and decisions to humans to analyse, categorise and tag. They will then be fed back into DoJo to continue the monstrous generation of data needed to take an AI child and turn it into an AI adult, capable of driving your vehicle.
Even today we are seeing that multiple paths are being stored locally and then used on a second, or even third, run without need for the software to be refreshed. The local system can then store this "known" decision path and ditch the others. It will then be forwarded to the central system as the correct way to handle the specific situation.
This is only going to get faster. Clearly Beta 8.1 hit a speed bump where data in DoJo plus AI software tuning met a sweet spot where enough edge cases had been conquered for the software to be rolled out to a bigger audience.
There is only one reason for needing that bigger audience. DATA. And they are going to get it. Reams of it and DoJo is going to eat it and feed it into the AI training.
If we think machine learning speeds, once FSD beta x.x has enough learning data to consume. Then, quite literally, it could move from Beta to R1.0 in a matter of days.
When that is done, Tesla will have billions of miles of actual real life driving where users did not have to intervene and in all conditions all over the US, sight unseen if need be. Which will be overwhelming evidence to allow certification at higher levels of autonomy, 3/4/5.
Then we can talk about whether Waymo geofenced "driverless" services, which give you the service they are capable of giving, within the area they are capable of giving it in, is really level 4 or is more of a level 3.5.
There is no doubt that when Tesla FSD finally comes out of Beta it will truly be Level 4 and, if not Level5, will very rapidly reach the point of being level 5.
FSD being Level 2, today, is nothing more than legality and politics! It has nothing at all to do with how good or capable FSD is, nor whether it will remain level 2 forever, which to me is nothing more than FUD.