Elon: FSD Beta tweets

[Mardak]

How will single stack improve parking lots

FSD Beta already drives in parking lots albeit (a lot) jerkily. My guess is that's because the training data has mostly been from actual city streets, and FSD Beta happens to recognize common features such as lines and curbs. My assumption is that single stack means Tesla will intentionally train the current FSD Beta networks with parking lot data to improve the driving behavior in parking lots such that the existing parking lot neural networks used for Summon are thrown away resulting in a single stack that drives on city streets and parking lots.

Yes there will still need to be some Summon-specific code, e.g., remote control from the app, but the driving behavior will at that point effectively just turn on "normal" FSD Beta code to drive to the pinned destination -- just without a driver needed in the seat.

https://twitter.com/x/status/1416511905515048966

 

[dcesq06]

Yes, if they have extra resources. Otherwise I'd concentrate on top 5 or 10 issues fsd testers are reporting on.

Its possible Tesla has a separate set of devs working on single stack (esp. the NN side, if most of the bugs are on the planner side). But I'd never prioritize this over fixing critical bugs in city fsd. Fixing bugs in city FSD would help reduce chances of future accidents which could damage the long term plans of Tesla.

So if Tesla has had A, B, C, and D teams working separately over the past 2 years on different aspects of the FSD experience, you would recommend pulling B, C and D off of their dedicated projects to focus solely on what A is doing?

More to the point, unless you know what resources Tesla has available, how they are allocating such resources, and how they are prioritizing such resources, then it seems a bit unfair to state your assumptions (on what Tesla is prioritizing) as if the assumptions are fact when, in reality, the assumptions are really just idle conjecture.

 

[EVNow]

So if Tesla has had A, B, C, and D teams working separately over the past 2 years on different aspects of the FSD experience, you would recommend pulling B, C and D off of their dedicated projects to focus solely on what A is doing?

More to the point, unless you know what resources Tesla has available, how they are allocating such resources, and how they are prioritizing such resources, then it seems a bit unfair to state your assumptions (on what Tesla is prioritizing) as if the assumptions are fact when, in reality, the assumptions are really just idle conjecture.

I'm just using my 2 decades experience managing software teams.

There are a lot of reasons for pulling people from one to the other team. It happens ALL the time. For eg. do you think people doing NOA two years back are all still doing NOA ? No - they would have been pulled to do FSD, leaving a skeletal staff there to provide any bug fixes needed. Teams are fluid - they are not fixed like bureaucracies.

Also, you want people to cross train. Good for the company and good / more interesting for the dev.

Remember they pulled solar city engineers to help with Model 3 production. Do you think they won't pull NOA, summon devs to do FSD ?!

 

[omarsultan]

Exactly my point!

It was pulled back from the refresh S too--I had it for a couple of FW releases, then it went away,

 

[EVNow]

My assumption is that single stack means Tesla will intentionally train the current FSD Beta networks with parking lot data to improve the driving behavior in parking lots such that the existing parking lot neural networks used for Summon are thrown away resulting in a single stack that drives on city streets and parking lots.

But single stack is not "free". Tesla needs to do a lot of work to make it happen and you should expect a lot of regressions.

If what you want is better summon - seems to me better to ask for better summon than single stack.

Single stack will remain the goal for Tesla - its about optimizing nodes and they need to do it at some point. But its more behind the scenes improvement rather than anything revolutionary for us "freely". They would still need to do the hard work of retraining fsd beta NN for parking lot.

 

[EVNow]

It was pulled back from the refresh S too--I had it for a couple of FW releases, then it went away,

Did it work ok ?

 

[omarsultan]

Did it work ok ?

Yeah, it was pretty cool--it was nice that it would pick spots without them being needed to be framed by other cars. It struggled a bit when lighting was less than ideal and the UI could use bit more tweaking.

 

[helvio]

Yeah, it was pretty cool--it was nice that it would pick spots without them being needed to be framed by other cars. It struggled a bit when lighting was less than ideal and the UI could use bit more tweaking.

On that note, what is your honest opinion about the Yoke vs a Wheel? I'm considering replacing mine with a Yoke to get the extra visibility.

 

[EVNow]

Yeah, it was pretty cool--it was nice that it would pick spots without them being needed to be framed by other cars. It struggled a bit when lighting was less than ideal and the UI could use bit more tweaking.

Wonder why they removed it.

Pre-Covid when I was commuting, I'd use autopark daily. Once you know the limitations, its very effective - though far from elegant.

 

[omarsultan]

On that note, what is your honest opinion about the Yoke vs a Wheel? I'm considering replacing mine with a Yoke to get the extra visibility.

I like it--I would not go back to a round wheel and driving my wife's X these days feels awkward these days. Bear in mind I went through a progression:

I am gonna die > OK, I can do this > no worse than wheel + stalk > better than wheels + stalk

What I really like is having key control under my thumbs--don't need to move my hands to do basic/everyday things. The best analogy I have it that is has similar ergonomics to an Xbox or PS controller: you keep a firm grip on the controller all the time and most things you need to do are under your thumbs.

The haptics still need some work, but they seem to be working on refining them--they recently released an update that is a marked improvement.

 

[Mardak]

But single stack is not "free". Tesla needs to do a lot of work to make it happen and you should expect a lot of regressions.

Yes it takes time to create triggers to collect video, label training data (even automatically), train the networks, run internal tests, deploy in shadow mode to evaluate, etc. The previous tweet I linked about "sublime" Summon was from July 17, so over 3 months ago, but that's probably a reasonable estimate of when Tesla was actively working on it. Elon Musk also reiterated and confirmed your "not free" with the comment about "massive NN retraining:"

https://twitter.com/x/status/1429903213726093315

It's less clear about "a lot of regressions" as clearly Tesla has not released it yet probably due to finding regressions. I wonder if this is part of the reason for the repeated delays of wider rollout as introducing regressions "steps back" to a larger audience could be more dangerous for those who don't follow as closely:

https://twitter.com/x/status/1364491362100477952

If what you want is better summon - seems to me better to ask for better summon than single stack.

As you probably know from managing software teams, it's sometimes better to reallocate a whole team instead of having them maintain a soon-to-be obsolete feature. Yes, there is uncertainty of whether the adjacent feature (FSD Beta) can fully replace the other (Smart Summon) especially in the near term, but it doesn't seem that unreasonable that the long term plan is for FSD Beta stack to handle "everything."

It sounds like you believe merging Summon into FSD Beta networks will require additions/modifications to the FSD Beta network architecture, e.g., new heads/outputs. I suppose there's parking lines, but those can appear on city streets, and at least from this old single main camera Autopilot view, that's likely something FSD Beta networks already predict (just not visualized):

 

[EVNow]

As you probably know from managing software teams, it's sometimes better to reallocate a whole team instead of having them maintain a soon-to-be obsolete feature. Yes, there is uncertainty of whether the adjacent feature (FSD Beta) can fully replace the other (Smart Summon) especially in the near term, but it doesn't seem that unreasonable that the long term plan is for FSD Beta stack to handle "everything."

Usually in such situations - where you have one project "winding down" but in production and the other ramping up - you move most people, but leave one or two people to manage bug fixes. The other option is to merge the teams and let the team internally manage bug fixes.

It sounds like you believe merging Summon into FSD Beta networks will require additions/modifications to the FSD Beta network architecture, e.g., new heads/outputs. I suppose there's parking lines, but those can appear on city streets, and at least from this old single main camera Autopilot view, that's likely something FSD Beta networks already predict (just not visualized):

Yes - I do think they will need specific new outputs that summon / parking will use. More importantly - they can't just use the old training data, if they are using videos to train the new fsd beta NN. They may need new data.

Finally, they will have to probably substantially change the planning code to use new NN.

While all this is needed to be done in the medium/long term, the team is probably under a lot of pressure to quickly improve FSD Beta with a ton of issues they are probably now getting. I expect them to concentrate on this in the short term.

 

[helvio]

While all this is needed to be done in the medium/long term, the team is probably under a lot of pressure to quickly improve FSD Beta with a ton of issues they are probably now getting. I expect them to concentrate on this in the short term.

I for one would be ecstatic if I could drive to work (on FSD Beta) and leave the car to let it park itself. It's perfectly legal given the parking lot is a private road and would be a big Wow from Tesla. Maybe they want to do that to keep the flame alive because they know the impact it would cause. Just saying, there's more than the critical bugs to fix. Image is just as important as functionality.

 

[EVNow]

Maybe they want to do that to keep the flame alive because they know the impact it would cause.

Imagine if you can actually drive to your work using FSD Beta - hmm ... without any "impacts"

 

[helvio]

Imagine if you can actually drive to your work using FSD Beta - hmm ... without any "impacts"

hahaah well played

 

[2101Guy]

I like it--I would not go back to a round wheel and driving my wife's X these days feels awkward these days. Bear in mind I went through a progression:

I am gonna die > OK, I can do this > no worse than wheel + stalk > better than wheels + stalk

What I really like is having key control under my thumbs--don't need to move my hands to do basic/everyday things. The best analogy I have it that is has similar ergonomics to an Xbox or PS controller: you keep a firm grip on the controller all the time and most things you need to do are under your thumbs.

The haptics still need some work, but they seem to be working on refining them--they recently released an update that is a marked improvement.

still another positive yoke review, but Im still not hearing from anyone "removing ability to activate horn via airbag was a step forward".

 

[Terminator857]

Elon confirms Tesla is eager to kick off people who are inattentive from beta.

https://twitter.com/x/status/1451240031604125704

 

[Dan D.]

Elon confirms Tesla is eager to kick off people who are inattentive from beta.

https://twitter.com/x/status/1451240031604125704

So only "While using FSD Beta". If you turn FSD Beta off and use straight AP and get strikeouts that won't get you kicked out.

 

[tdiggity]

So only "While using FSD Beta". If you turn FSD Beta off and use straight AP and get strikeouts that won't get you kicked out.

Don't be so sure they have separate reporting for this But, at least you get a warning!

 

[diplomat33]

https://twitter.com/x/status/1451354738935615489