Elon: FSD Beta tweets

[Mardak]

That was also right around the month that FSD TRANSFER started, could have been an uptick in new to a new platform users?

I tried estimating the monthly mileage based on actual dates from quarterly reports and simple extrapolation from the partial April 2024 numbers:

Original FSD Transfer and price drop to $12k were both September 2023, and that does seem to have resulted in nearly 10% boost in monthly miles. The large jump from April 2023 was wide release of single stack 11.x including highways, and the large jump now is from 12.x and free trial.

At least 2 years ago, Elon Musk said success needs 10+ billion miles of vehicle data, but presumably that's not necessarily with FSD active as training data from shadow mode data collection from the whole fleet accumulates billions of miles much faster?

https://twitter.com/x/status/1503888110899376138

 

[AlanSubie4Life]

The large jump from April 2023 was wide release of single stack 11.x, and the large jump now is from 12.x and free trial.

So if I am understanding correctly, basically we are seeing a bunch of freeway miles, as those AP miles (using the old Autosteer/NOA stack previously) are recategorized as “FSD” (Supervised) miles, as people switch to FSD (v11) on the freeways from Autosteer (Beta). Similar to the big step last year when people in the FSD Beta program started using FSD v11 single stack on the freeway, resulting in a lot more FSD miles. Same thing. Just a bunch more users (2.5-3x) now who are doing the same thing without paying.

So not nearly as much of an increase in FSDS use (of the type people think is being displayed here) as one would think, most likely.

Going to be sad when the curve flattens out again (meaning lower slope, reverting to close to prior slope).

I wonder when they will start displaying City Streets miles?

Of course your projections may be right - it is possible that Tesla will just start giving FSD to the full fleet indefinitely and stop charging for it (it is not likely a significant source of revenue for them). After all, Elon said in previous Tweets that safety features would always be free. If it’s safer it is no longer a convenience feature!

 

[AlanSubie4Life]

From Green:
According to Tesla personnel here on twitter, the minuscule hw changes on legacy s/x (different backup cam) are so insurmountable, Tesla cannot even fully validate it in reasonable time.

https://twitter.com/x/status/1777772133088624817

If we declare legacy S/X HW changes as big enough to impede validation of adas for a long time, CT is like alien technology in comparison,...

https://twitter.com/x/status/1777773765608325577

They would also have to have the CT just drive over the median since even the Model Y doesn’t fit apparently. That will need different training.

 

[Ben W]

[…] L4 supervised […]

 

[diplomat33]

View attachment 1037089

That is a common misconception that if the system is supervised, it is L2. L4 supervised is not L2. Read the SAE levels again.

 

[azred]

Keep in mind Autonomy day was March 2019 and they were sure they were close then too

And of course there was the prediction in early 2017 that a totally autonomous drive from CA to NY would happen in December 2017. When ranking Elon’s whoppers that was probably the all-time lie…so far, but give him time. His August Robotaxi intro will surely include some totally outrageous claims.

 

[azred]

I tried estimating the monthly mileage based on actual dates from quarterly reports and simple extrapolation from the partial April 2024 numbers:

Original FSD Transfer and price drop to $12k were both September 2023, and that does seem to have resulted in nearly 10% boost in monthly miles. The large jump from April 2023 was wide release of single stack 11.x including highways, and the large jump now is from 12.x and free trial.

At least 2 years ago, Elon Musk said success needs 10+ billion miles of vehicle data, but presumably that's not necessarily with FSD active as training data from shadow mode data collection from the whole fleet accumulates billions of miles much faster?

https://twitter.com/x/status/1503888110899376138

Nope, Sept 2023 was not the start of the first FSD transfer offer. We did our transfer deal in July 2023. Perhaps Sept 30 was the last day of the offer?

 

[AlanSubie4Life]

Nope, Sept 2023 was not the start of the first FSD transfer offer. We did our transfer deal in July 2023. Perhaps Sept 30 was the last day of the offer?

It doesn’t really matter, it is a fairly inconsequential impact anyway. Just kind of some noise in the incremental miles in that timeframe.

I would just focus on the big steps. The really huge ones are just highway Autopilot miles getting added into the total when they weren’t being counted before.

Just an accounting gimmick causing two large increases in slope, because technically it started using FSD for those miles.

It’s not like people suddenly started pulling down on the stalk 3 times as much in April 2023. And even if the number of FSD City Streets miles being driven went up a factor of 3 in early April, it is not possible to say how many miles that was.

 

[Ben W]

That is a common misconception that if the system is supervised, it is L2. L4 supervised is not L2. Read the SAE levels again.

It’s a mixture of L2 (in the “what does the human driver have to do” column) and L4 (in the “features” column). It could just as easily be called “Universal L2” instead of “Supervised L4”,

My main objection to labeling it “[qualifier] L4” is that I don’t believe the current generation of hardware (HW3/HW4) will ever be fundamentally capable of safe unsupervised city-streets operation. It may certainly get a couple orders of magnitude better than it is now, but not (from my vantage point) the 4-5 orders of magnitude necessary for UNsupervised L4, until HW5/HW6.

Acknowledged that Tesla is using their own SAE-independent term for it (“Full Self-Driving (Supervised)”); I’m not sure whether it’s directly referred to as L4 in Tesla’s material? But in general I think they would be better to under-promise and over-deliver, than the other way round.

 

[JulienW]

It’s a mixture of L2 (in the “what does the human driver have to do” column) and L4 (in the “features” column). It could just as easily be called “Universal L2” instead of “Supervised L4”,

My main objection to labeling it “[qualifier] L4” is that I don’t believe the current generation of hardware (HW3/HW4) will ever be fundamentally capable of safe unsupervised city-streets operation. It may certainly get a couple orders of magnitude better than it is now, but not (from my vantage point) the 4-5 orders of magnitude necessary for UNsupervised L4, until HW5/HW6.

Acknowledged that Tesla is using their own SAE-independent term for it (“Full Self-Driving (Supervised)”); I’m not sure whether it’s directly referred to as L4 in Tesla’s material? But in general I think they would be better to under-promise and over-deliver, than the other way round.

It is 100% L2 and NOTHING else.

 

[Knightshade]

It’s a mixture of L2 (in the “what does the human driver have to do” column) and L4 (in the “features” column).

There's no such thing as "partially L4"

It's an L2 system because it requires a human at all times, period full stop.

Tesla is making up their own terms because they want to make it sound more advanced/capable than the SAE level (which is 2) would suggest.

 

[Dewg]

They would also have to have the CT just drive over the median since even the Model Y doesn’t fit apparently. That will need different training.

View attachment 1037074

This is starting to look like the stop short issue, at stop signs and red lights, it stops several feet from the line. With the median it's doing the same thing, stopping short of the artificial line it created in the median.

 

[AlanSubie4Life]

This is starting to look like the stop short issue, at stop signs and red lights, it stops several feet from the line. With the median it's doing the same thing, stopping short of the artificial line it created in the median.

I am pretty sure the Cybertruck will not fit regardless. Though maybe they’ll use the front camera for perfect positioning. As long as they don’t do any ridiculous looping-right-to-turn-left nonsense.

Anyway, I think using the median is dumb in this situation and they should just always plan to roll it. Median only makes sense if it is all driveable median, and most of the time unnecessary even then and puts the vehicle at greater risk.

Fortunately for Cybertruck it is all driveable. Just merge in from the grass.

 

[Ben W]

It is 100% L2 and NOTHING else.

View attachment 1037121

I believe this interpretation is not correct. SAE has since clarified that L2 “steering” simply means “lanekeeping”: https://www.sae.org/binaries/content/assets/cm/content/blog/sae-j3016-visual-chart_5.3.21.pdf

Tesla’s current features (e.g. merging, turning, following traffic signals, point-to-point navigation) are clearly more advanced than that, somewhere between L3 and L4 in the “example features” column. Though agreed that in the “what the driver must do” column they are still squarely L2.

 

[willow_hiller]

Can we please not argue the definition of a standard, again?

Standards, including the SAE levels, are written with an intent that there be one correct interpretation. There is no point in suggesting alternative interpretations.

We could come up with a separate definition for levels of autonomy based on capabilities instead of liability and driver monitoring that more closely follows Tesla's development path. That would be a way more productive exercise than trying to reinterpret the SAE levels.

 

[Ben W]

There's no such thing as "partially L4"

It's an L2 system because it requires a human at all times, period full stop.

Tesla is making up their own terms because they want to make it sound more advanced/capable than the SAE level (which is 2) would suggest.

Agreed that a system must meet all the requirements for L4 (in every column) to be labeled “L4”. My statement was just that in the “Features” column they are already well past L2, somewhere between L3 and L4. I’m not saying this entitles them to call themselves an “L4 system”. (Agreed they are not.)

 

[Ben W]

I believe this interpretation is not correct. SAE has since clarified that L2 “steering” simply means “lanekeeping”: https://www.sae.org/binaries/content/assets/cm/content/blog/sae-j3016-visual-chart_5.3.21.pdf

Tesla’s current features (e.g. merging, turning, following traffic signals, point-to-point navigation) are clearly more advanced than that, somewhere between L3 and L4 in the “example features” column. Though agreed that in the “what the driver must do” column they are still squarely L2, which makes their system still squarely an L2 system.

 

[Ben W]

Standards, including the SAE levels, are written with an intent that there be one correct interpretation. There is no point in suggesting alternative interpretations.

When they use a highly ambiguous phrase (“steering”), it will be interpreted in different ways. That’s why they disambiguated it to “lanekeeping” in an updated graphic. Unfortunately, the old graphic is still floating around.

We could come up with a separate definition for levels of autonomy based on capabilities instead of liability and driver monitoring that more closely follows Tesla's development path. That would be a way more productive exercise than trying to reinterpret the SAE levels.

Agreed that the SAE system implies lockstep progress in different categories that doesn’t reflect real-world development. Tesla’s system is what it is, and doesn’t exactly fit any of the SAE vertical slices, which is why Tesla is giving it its own label and terminology.

 

[Knightshade]

I believe this interpretation is not correct. SAE has since clarified that L2 “steering” simply means “lanekeeping”

Yeah, no, they have not. You really need to actually read J3016. Specifically where they explain what lateral vehicle motion control is- which is what they list on the chart as steering.

LATERAL VEHICLE MOTION CONTROL
The DDT subtask comprising the activities necessary for the real-time, sustained regulation of the y-axis component of
vehicle motion

A right turn at an intersection would ALSO be lateral vehicle motion control. AKA steering.

At no point, ever, did they define it simply as "lanekeeping"-- though they do give that as one example of a task among the things that are included in lateral vehicle motion control-- and in the doc they note such control includes detection of the vehicles position relatively to lane boundaries. Including a specific thing inside the definition is not the whole definition.

Likewise brake/acceleration support on the simplified chart is more explicitly defined as:

3.15 LONGITUDINAL VEHICLE MOTION CONTROL
The DDT subtask comprising the activities necessary for the real-time, sustained regulation of the x-axis component of
vehicle motion

L1 does one of those, and always requires a human to be present and engaged in the DDT. L2 does both, and always requires a human to be present and engaged in the DDT.

A system that does both REALLY well in REALLY complex situations, but still always requires a human to be present and engaged in the DDT is L2, period full stop.

Agreed that the SAE system implies lockstep progress in different categories that doesn’t reflect real-world development. Tesla’s system is what it is, and doesn’t exactly fit any of the SAE vertical slices, which is why Tesla is giving it its own label and terminology.

it does though.

It fits in exactly L2, because it always requires a human driving the car. 100% of the time.

You seem to be trying to slice up the chart and remix it, which isn't how the levels work at all. Perhaps instead of just looking at the chart you should read the detailed descriptions in J3016 to understand why the levels are how they are, and why Teslas system is categorically not "somewhere between L3 and L4 " as you claimed.

There is no such thing as "between" levels.

Either you meet all the criteria for a level, or your system is at a lower level where it DOES meet all the criteria.

For Teslas current FSD, that's L2.

 

[Ben W]

Yeah, no, they have not.
At no point, ever, did they define it simply as "lanekeeping”.

Ah, I see how I misread the chart thinking “lanekeeping” was part of the definition, rather than merely an example. I suppose I have to walk back some of my recent posts.

Still, it’s understandable that since “L2” encompasses such a huge range of implementations (from “adaptive cruise control + lanekeeping“ to “supervised point-to-point chauffeur”), it’s understandable why Tesla wouldn’t want to call their system simply “L2” and have it potentially confused with the more basic end of the spectrum.