FSD rewrite will go out on Oct 20 to limited beta

[mark95476]

Makes me wonder if this is friendly competition or a beef?

SpaceX Starlink more or less killed Google's Loon moonshot. Fortunately, Google Fiber plays in different markets.

I do not know how, but I missed this Elon reply.....
View attachment 630554

src: https://twitter.com/elonmusk/status/1353564655088619520

 

[mhan00]

Along with a quantum leap in the number of disengagements per mile

I mean, that’s just totally expected and natural. The car got massively more capable, which means it can handle far more situations, and thus there are more situations where it can screw up. Increased disengagements are no surprise and entirely natural. It’s very impressive what the cars can do right now, even with the disengagements and screw ups.

 

[heltok]

New system also means new bugs. As they iron out low hanging fruit bugs disengagements rapidly decreases.

I really look forward to FSD wide release so that Waymo, Mobileye, legacy industry, bears etc can get their wake up call and realize how wrong they have been on. What Tesla are doing in production today is so far ahead of where competition hopes to be in production in a few years. While doing it in a low cost way, in a car made to last 1M miles, in a car with low cost per mile etc.

 

[Mardak]

who cares how the revisions are numbered? I'm sure the change makes sense to the company.

If people are wondering why Elon Musk had to spend any effort at all about something as trivial as version numbering for software that's used by a small beta audience. Most likely it's because he commented about expanding the beta around version 10 or 11: https://twitter.com/elonmusk/status/1342909163903455233

But his intent was counting FSD driving improvements as that is actually what makes progress in getting FSD ready for a wide release. Elon tweeted about a "beta 5" and the community just added 1 to the beta number for every release after that including the holiday update, which had no AP changes. But then someone at Tesla followed along with the numbering and put in "10.1" into the software displayed version trying to indicate there was no major changes, so Elon likely came in to reset things to how he wanted.

    1: 2020.40.8.10
    2: 2020.40.8.11
    3: 2020.40.8.12
  3.1: 2020.40.8.13 (non-AP bug fix)
    4: 2020.44.10.2
    5: 2020.44.15.3 ("beta 5")
    6: 2020.44.15.4
    7: 2020.48.10.1
    8: 2020.48.12.15
    9: 2020.48.26.1 (holiday update)
   10: 2020.48.35.1
 10.1: 2020.48.35.6 (restore expanded view)
  8.1: 2020.48.35.7 (renumber versioning)

 

[Bladerskb]

Welp, I guess this means the competition is deader than dead.

Deader than dead by using mapped intersection, mapped lanes, mapped direction of travel, mapped traffic lights and stop signs and traffic light lane relevancy, stop lines, and more. Near 100% hand written conventional c++ driving policy and more importantly, mapped drive-able path.

You people claim that Waymo and others are like trains because they follow a path that is laid out like a train. That's literally what Tesla does. They use crowdsourced drive-able path, just another thing they copied from Mobileye. They have the path of every lane in an intersection. This is why the system struggles [even in simple situations] the minute they get to a location where not alot of Teslas have been in, hence not fully mapped.

As @verygreen said: "There's no actual intelligence in the system."

Mapped drive-able path. FSD Beta is basically a glorified path following train system.

Early Mobileye REM test

 

[powertoold]

Early Mobileye REM test

Mobileye is dead to me. A joke at best.

They don't have what it takes to achieve FSD, sorry.

 

[willow_hiller]

It's really fun that I can now use the things you said about Waymo to make it look like you're arguing with yourself:

Mapped drive-able path. FSD Beta is basically a glorified path following train system.

So then its not a equivalent to a train. Train moves in one direction. They have no obstacles. There's are no negotiations with other real time road users. They don't squeeze through, go around something or someone or cut through.
This comparison is absolute nonsense and shows the absurd logic tesla fans will attempt to use to downplay anyone else. This is why i don't participate in these discussions anymore. Its pointless.

 

[Bladerskb]

It's really fun that I can now use the things you said about Waymo to make it look like you're arguing with yourself:

Correct I'm using what Tesla fanboys are saying negatively about Waymo, Mobileye and others on all social media and using the same line of thinking to analyze how Tesla's FSD Beta operates. Turns out it literally operates the exact same way. As verygreen said, they mindlessly follow the path laid out and only avoid obstacles. When they get to a city with no mapped drivable path due to less Teslas being driven there, they struggle for the simplest thing.

Hence I said "You people claim that Waymo and others are like trains because".

If these Tesla fans weren't simply misinformed fanboys who refuse to think for themselves, they will admit that either they were wrong or will apply their same thinking to Tesla and call FSD Beta a glorified train as @verygreen eluded to.

But no they won't. They will come up with another excuse.

Proving there's absolute not one iota of logic, fact or reason to anything these people say.
Its straight Myth and fairy tale.

 

[willow_hiller]

Turns out it literally operates the exact same way.

Cool, so when will Waymo give their vehicles to 1,000 individuals all across the US to use in any location what-so-ever, regardless of whether a Waymo vehicle has driven that route before?

 

[Bladerskb]

Cool, so when will Waymo give their vehicles to 1,000 individuals all across the US to use in any location what-so-ever, regardless of whether a Tesla vehicle has driven that route before?

They already have, they are testing in 24 cities. They don't need Youtube celebrities to market and PR for them.

When will Tesla have a car in any city that you can sleep in while it drives you?

Your god Elon is using mapped driveable path just like Waymo and Mobileye and still can't get past one disengagement every 10 or so miles. That's pathetic.

 

[willow_hiller]

They already have, they are testing in 24 cities. They don't need Youtube celebrities to market and PR for them.

When will Tesla have a car in any city that you can sleep in while it drives you?

Your god Elon is using mapped driveable path just like Waymo and Mobileye and still can't get past one disengagement every 10 or so miles. That's pathetic.

Care to provide a source for that disengagement statistic, or did you make it up after watching Youtube celebrities?

 

[Bladerskb]

Mobileye is dead to me. A joke at best.

They don't have what it takes to achieve FSD, sorry.

Tesla is dead to me. A joke at best.

Recent FSD videos, their heavy use of hd maps, crowdsourced driveable-path sand 100% hand written driving policy proves they don't have what it takes to achieve FSD, sorry.

I thought they had a lottery chance, now I see they are hopeless. If all they are gonna do is copy mobileye.
I'm willing to bet on that with anyone, just like I was willing to back in 2017.Though there was alot of big-mouth people like you, no one put their money where their mouth are. Because they know they are full of it.

 

[powertoold]

A joke at best.

Mobileye, especially Amnon, is the biggest joke in FSD, my opinion:

One of the many stupid in this video: "regulators will want to see the parameters in the RSS system" LOL (I'm paraphrasing)

 

[Mardak]

One thing people have often gotten confused about neural networks is believing they're like a database with memorized behaviors for a given situation. This was discussed 5 years ago with AlphaGo during the Go matches against Lee Sedol where the AI showed superhuman understanding, and more recently with AlphaZero Chess matches against Stockfish with complaints of unfair advantages of "opening book" which traditional chess engines use without really needing to "think" early in the game.

Instead the neural network tries to generalize with a variety of training data leading to some behaviors that have higher probabilities than others. This generalization is what Tesla is betting on for Autopilot to handle situations that a Tesla vehicle has never driven before. Just like in Go and Chess where there are so many potential positions that practically one couldn't store all possible positions and the correct behavior, FSD beta neural networks (and even if you include map data) don't store what to do for any given road let alone all possible driveable spaces.

 

[Bladerskb]

One thing people have often gotten confused about neural networks is believing they're like a database with memorized behaviors for a given situation. This was discussed 5 years ago with AlphaGo during the Go matches against Lee Sedol where the AI showed superhuman understanding, and more recently with AlphaZero Chess matches against Stockfish with complaints of unfair advantages of "opening book" which traditional chess engines use without really needing to "think" early in the game.

Instead the neural network tries to generalize with a variety of training data leading to some behaviors that have higher probabilities than others. This generalization is what Tesla is betting on for Autopilot to handle situations that a Tesla vehicle has never driven before. Just like in Go and Chess where there are so many potential positions that practically one couldn't store all possible positions and the correct behavior, FSD beta neural networks (and even if you include map data) don't store what to do for any given road let alone all possible driveable spaces.

STOP SPREADING MISINFORMATION FOR CRYING OUT LOUD!!!!

This is not how any of this work.
This is exactly what am talking about about myths and misinformation.
Its staggering. You people spread misinformation like wildfire. If you don't know what you're are talking about.
Then don't post. Seriously STOP!

AlphaGo and AlphaZero uses reinforcement learning networks while Tesla FSD uses near 100% hand written c++ conventional algorithm.
Which part of that don't you people understand?

FSD Beta driving policy consists of codes like A* algorithms and other conventional control algorithms and c++ code makes all the decisions.:
blog.Habrador.com: Explaining the Hybrid A Star pathfinding algorithm for selfdriving cars

Tesla's FSD Beta ISNT. IS NOT anyway like AlphaGo or AlphaZero.

Just like in Go and Chess where there are so many potential positions that practically one couldn't store all possible positions and the correct behavior, FSD beta neural networks (and even if you include map data) don't store what to do for any given road let alone all possible driveable spaces.

NO. FSD Beta stores exactly what to do for every given road. That what the drivable path is there for. The car literally just follows it. Every single road has a drive-able path. Intersections that don't are obvious and FSD Beta struggles terribly at them.

As @verygreen calls it. Blind mimicking.

 

[ZeApelido]

Talking about what a system currently does (C++ driving policy), and assuming that's what it's going to do in the future, is tenuous.

 

[powertoold]

FSD beta / AP traffic control don't blindly mimick. That's completely wrong.

Tesla uses maps as a guide, once they're at the location, the perception will confirm the presence of traffic controls / lanes / road markings / etc. Karpathy talked about this during his description of the stop sign feature.

Starts 27:00 here:

 

[Bladerskb]

Talking about what a system currently does (C++ driving policy), and assuming that's what it's going to do in the future, is tenuous.

Always with the future when it comes to Tesla.
Just you wait, it will be sentient soon and just after that it will turn into a time machine.
Lets ignore the facts at hand.

FSD beta / AP traffic control don't blindly mimick. That's completely wrong.

Tesla uses maps as a guide, once they're at the location, the perception will confirm the presence of traffic controls / lanes / road markings / etc. Karpathy talked about this during his description of the stop sign feature.

Starts 27:00 here:

Its not just a guide it follows the driveable path laid out before. Without a driveable path it fails miserably.
Not only that, but FSD Beta will stop without the presence of traffic control as long as the map has a traffic control (lights or stop sign).

I could go on and on.
Hence the wording 'blindly mimicking' as verygreen said.
A glorified path following train.

 

[powertoold]

Hence the wording 'blindly mimicking' as verygreen said

Nope, that's wrong. It's not blindly mimicking the path. Otherwise, it wouldn't be able to avoid obstacles, duh.

 

[Mardak]

Talking about what a system currently does (C++ driving policy), and assuming that's what it's going to do in the future, is tenuous.

One would think there will always be some traditional C++ "software 1.0" style code even many iterations later as that code is what passes data into the neural network to evaluate and get outputs. AlphaZero style AIs have traditional code to aggregate neural network value and policy predictions to do even better by having the network evaluate potential future positions. Similarly FSD beta uses traditional code to decide which predictions to use, e.g., pick a path towards the right for an upcoming turn, or ignore, e.g., don't drive through a traffic calming island even though the network believes driving straight is the most common behavior.

MuZero is AlphaZero's successor and reduced even more of the traditional software that would have provided truth of what can happen in potential future states (e.g., whether a move is allowed), and there the neural network became much larger and learned to predict how various games (Go, Chess, Atari) would behave when doing certain actions. But even then, there was still traditional code to decide how to aggregate the various predictions and do the decided action, e.g., move a piece or steer or accelerate. People have speculated more "end-to-end" training of Autopilot neural networks where the neural network gets sensor and other inputs and directly decides steering and acceleration, and that presumably would replace a lot of traditional software, but that is most likely quite some time away.

From watching FSD beta videos, it's pretty clear to me that there's plenty of of traditional code for current & upcoming lane selection behavior, but how it gets into or "stays" in a lane (e.g., avoiding objects) is primarily neural network predictions as suggested by the wobbly path visualization.

One would expect Autopilot to move more of the lane selection behavior into the neural network especially in situations that are misbehaving due to incorrect map data, but indeed we have no insights into when that would happen and even when it does, how much is still left as traditional code.