FSD rewrite will go out on Oct 20 to limited beta

[CurtisVL]

By your logic the vNew FSD would lead to the same behaviors as the vOld FSD. An existing, tested model executes an existing, tested action. If either of those statements is false, either the model or the action being untested, then by default the chain is untested.

You're absolutely right that the model will benefit in object recognition by all the billions of miles of real world data. That model will be further enhanced by the ability to recognize and track objects across multiple camera views, as Elon's said publicly. That improved model has not been tested by billions of real world miles.

Similarly, you have logic executed based on those models. Today, you will see phantom breaks due to any number of interpreted signals from that model. Elon has suggested that no longer happens. That new, changed behavior has not been tested by billions of real world miles.

This is in contrast to today, where adding one additional piece of functionality doesn't invalidate the entirety of AP, because the change is supplemental and iterative, vs. a fundamental replacement.

If you work in software, which it sounds like you do, you know that the con of a rewrite is that it has the potential to introduce far more bugs than a single, incremental update to an existing codebase.

Though, having said that, vOld FSD was at a point where training wasn't making massive changes. It's very likely the re-write will release in a funky state and in the coming weeks/months it'll rapidly improve, especially with the additional data it'll be working with due to using all cameras instead of just 5, and with the '4th dimension' (time) for a bit of context to what its seeing.

Some issues will be fixed for sure, for example, long vehicles such as trucks (semi's) can cause issues for the car as they switch from one camera to another, the car confuses the depth so the position flickers around a lot.

Regardless, I'm excited to see what's changed! I do expect the first times we're shown it that it'll be pretty buggy, but it should improve pretty fast once it starts getting worldwide data rather than just parts of California data mainly.

 

[Spaceosaurus]

This is true, however as you probably know 75% of modern software development is about testing frameworks, writing test scenarios and collecting/labeling test data. The actual code is only 25% of the work.

Done correctly you can do fairly substantial rewrites, and validate it against your testing framework and it should work quite well out of the box. Eventual loopholes found during early testing is used to improve the tests.

Even though Tesla doesn't collect billions km of driving, they've hopefully built an impressive automated testing suite from 4 years of production experience.

Oh I'm sure they do (or at least I hope they do ), but of course it's very very difficult to have a truly comprehensive test suite. That's why most companies have moved away from huge test suites and QA and into experimentation in production. You'll always have automation for your P0 base cases, but getting every edge case is super hard.

I imagine Tesla has an even harder time because both roads and other drivers don't have to play by the rules.

 

[Ryan H]

I wonder ...if fsd has been redeveloped from ground up....what does this mean to those who did not buy fsd? In otherwords, is tesla going to maintain a 'have' and 'have nots' version? (the resources..yikes) ...or will it be like windows 'home' vs 'pro' version? Personally I can't wait since I have purchase fsd, but nervously excited that we're all lab rats...hopefully not perceived as disposable.

Long term? I really doubt it will be separate. While the limited beta testing is going on, I think separate is the only way.

 

[Snuffysasa]

I think this could be a requirement for beta testers:
Tesla hacker reveals what driver-facing camera is looking for - Electrek

Possible, but I don't think so

 

[Mardak]

if fsd has been redeveloped from ground up....what does this mean to those who did not buy fsd?

The new software has been written targeting the newest and most common hardware, i.e., FSD computer / HW3 is in more than 50% of the fleet. Given that Tesla reports vehicle safety numbers for active safety, the new software will most likely deployed to most vehicles to get the safety benefits of the rewrite. So assuming that, Basic Autopilot features, e.g., lane keeping, and Enhanced Autopilot features, e.g., smart summon, will see benefits as well.

 

[RTPEV]

I wonder ...if fsd has been redeveloped from ground up....what does this mean to those who did not buy fsd? In otherwords, is tesla going to maintain a 'have' and 'have nots' version?

I don't follow what you're saying. Why would people that did not buy FSD receive a different version (or any version*) of FSD? And even if they did, why would that matter anyway, given that they don't have FSD?

* Okay, we do need to consider the fact that in reality, FSD is running in shadow mode on cars that do not have FSD, so I would expect that all vehicles capable of running the rewritten FSD NN/software will likely get the update once it goes into wide release**

** And now we need to consider that pre-HW3 cars will not be able to run the latest FSD software. So in THAT sense, there may be a subset of vehicles that have not bought FSD that are running a subset of the code (or maybe even fork of the old (current) software. Is this what you are talking about? And if so, again, why is that significant? I guess you could say that future active safety features may work better on HW3 equipped vehicles, so there may be some "have-nots" on HW2, but you can also make the statement that whatever active safety features they have are probably going to exceed what the car had when they first bought it.

 

[nimrooz]

I don't follow what you're saying. Why would people that did not buy FSD receive a different version (or any version*) of FSD? And even if they did, why would that matter anyway, given that they don't have FSD?

* Okay, we do need to consider the fact that in reality, FSD is running in shadow mode on cars that do not have FSD, so I would expect that all vehicles capable of running the rewritten FSD NN/software will likely get the update once it goes into wide release**

** And now we need to consider that pre-HW3 cars will not be able to run the latest FSD software. So in THAT sense, there may be a subset of vehicles that have not bought FSD that are running a subset of the code (or maybe even fork of the old (current) software. Is this what you are talking about? And if so, again, why is that significant? I guess you could say that future active safety features may work better on HW3 equipped vehicles, so there may be some "have-nots" on HW2, but you can also make the statement that whatever active safety features they have are probably going to exceed what the car had when they first bought it.

I think he was talking about the different iterations that FSD have been through, or descriptions if you would like to call it that.
Like me, when I bought my car and FSD, the FSD description of what I was buying had far more functions than the FSD (description) you are buying now.

And also, you have the EAP, which does several (almost all) of the new FSD description...

 

[RTPEV]

I think he was talking about the different iterations that FSD have been through, or descriptions if you would like to call it that.
Like me, when I bought my car and FSD, the FSD description of what I was buying had far more functions than the FSD (description) you are buying now.

And also, you have the EAP, which does several (almost all) of the new FSD description...

Ooops, I need to fix that. I bought FSD when they offered the $2K upgrade.

I still think that when you buy FSD you are buying into what it will eventually become (whatever you believe that will be), not its current functionality or any specific list of functions that may be coming out "next". Am I mistaken?

 

[NHK X]

All I know is that I will be driving perfect speed limit and my hand will be completely on the wheel

 

[diplomat33]

Elon continues to hype up the FSD rewrite.

 

[NHK X]

Elon continues to hype up the FSD rewrite.

Vague language but hope it crushes expectations

 

[diplomat33]

Don't get me wrong. As a Model 3 owner who paid for FSD, I would love for the FSD rewrite to be real autonomous driving. But I am skeptical. Too many false promises and disappointments. After all, AP2 was supposed to be FSD capable and then Elon admitted they needed AP3. AP3 was supposed to be FSD capable, and then Elon was like "actually, we need a rewrite". Now, he is claiming they were stuck in a local max but this rewrite is the true FSD. Let's see next week.

 

[ChrML]

Oh I'm sure they do (or at least I hope they do ), but of course it's very very difficult to have a truly comprehensive test suite. That's why most companies have moved away from huge test suites and QA and into experimentation in production. You'll always have automation for your P0 base cases, but getting every edge case is super hard.

I imagine Tesla has an even harder time because both roads and other drivers don't have to play by the rules.

Cases are usually added as they occur in production. Which means the test suite covers more edge scenarios the longer it's tested in production, regardless if you're on codebase 1, 2 or 3. Hopefully every Nth thousand test scenario reproduced in simulation eventually add another 9 to the reliability %.

My impression is not that devs test less, but they are better these days on scaling their test environment without bloating it or making it too tight with the actual code.

For instance we have a large product that would take 6 months just to function test end-to-end. A good testing environment allows us to release updates 3 times a week with very little manual smoke testing. FSD is like the same times 100000 in number of edge scenarios, but same principle.

 

[Bladerskb]

Its funny how a rewrite from scratch is considered and hyped as a game changer and industry leading when in reality it just shows who have been doing it wrong all along and just now trying another tech approach. While the people (Waymo...etc) who already have driverless cars ferrying riders all over a city is considered 5 years behind.

 

[mswlogo]

I wonder if we'll see the "same" capabilities on the new rewrite first. Once that's stable they will gradually add FSD features over time.

 

[nimrooz]

I think he was talking about the different iterations that FSD have been through, or descriptions if you would like to call it that.
Like me, when I bought my car and FSD, the FSD description of what I was buying had far more functions than the FSD (description) you are buying now.

And also, you have the EAP, which does several (almost all) of the new FSD description...

Ooops, I need to fix that. I bought FSD when they offered the $2K upgrade.

I still think that when you buy FSD you are buying into what it will eventually become (whatever you believe that will be), not its current functionality or any specific list of functions that may be coming out "next". Am I mistaken?

Guess the last part I wrote could be misunderstood, I don't know what you have.

But when you buy a Tesla you have several options, depending on what is available when you buy.
Different FSD descriptions, EAP, etc

We will see what FSD will become, but I sincerely hope that it will be like the description I bought from rather than the later description which has been more or less a step back and more like EAP in description...

 

[RTPEV]

Guess the last part I wrote could be misunderstood, I don't know what you have.

Ah, gotcha. I thought you looked at my signature and were talking about me specifically.

But when you buy a Tesla you have several options, depending on what is available when you buy.
Different FSD descriptions, EAP, etc

FSD = Full Self Driving, which means that if you have it (regardless of when you bought it) will eventually have the ability to drive around all by itself without any humans inside and participate in the Tesla Network, earning you money.

The "descriptions" I think you are referring to are its current or estimated short term functionality. I.e. the ability to stop at stop signs and traffic lights, recognize speed limit signs, traverse through green lights, take turns on surface streets, etc. That does not mean that if you buy FSD now that you are only buying that limited set of features. You are buying the whole package up to and including the eventual fully autonomous operation, if you believe that Tesla will eventually reach that point, that is. Rest assured, if/when you bought FSD you bought ALL future upgrades, and that list is bound to be quite numerous, and in my opinion, a very length process.

 

[LowlyOilBurner]

“The rewrite is done, but, you will need more compute power to run it. HW4 will go into production Q4 of 2021”

-Elon, probably

 

[willow_hiller]

Its funny how a rewrite from scratch is considered and hyped as a game changer and industry leading when in reality it just shows who have been doing it wrong all along and just now trying another tech approach. While the people (Waymo...etc) who already have driverless cars ferrying riders all over a city is considered 5 years behind.

How many times has Waymo rewritten their firmware? If they're not reworking their methods, how much better can we expect Waymo's autonomy to get? How do we know they're not stuck in a local maximum, and they've achieved everything possible with their current system? When will Waymo expand outside of a few select cities?

 

[powertoold]

Tesla is the only company that can actually do a rewrite, deploy it to their fleet, and then regather all the necessary real world data / miles within a reasonable time frame. Other companies have to work with what they have since they've invested so much already into their simulations and contracted drivers.