Again, there's already a lot of known info on all these topics, but some folks here seem much more interested in being angry than informed.
See, great example of this in the very post I'm quoting here.
He asks for evidence of a specific claim I made being based on anything from Tesla.
I provide evidence direct from Elon Musk supporting that specific claim.
He gives the post with exactly what he asked for a disagree.
To him, being angry > being informed. I'd bet money he'll give this one a disagree for pointing it out too.
Again, hopefully this isn't everyone's take on it.
If it's not, here's another relevant bit of info on managing expectations-
That's from over a year ago, setting some expectations about how widely the program would expand in several future versions.
8.2, a week later, expanded the program by 1000 people (still only like 2000 at the time)
The 8.3 mentioned there ended up not going out at all, being replaced by 9.0, and would be 20,000 people per that tweet (10x)
We know by end of Jan 2022 they had again tripled that to ~60,000 on the 10.x branch- and all in the US.
Now, what do we know SINCE then?
We know they appear to have virtually ceased rollouts to brand new US users for the moment.... but they are pushing now to Canadian users for the first time.
Again this is consistent with Tesla wanting
new data- rather than just increasing the volume of data for areas they already have good training sets for.
Additionally, as noted elsewhere, the updates even for existing fleet are different in recent months-- instead of almost the entire test fleet getting wide rollout of new versions within a few days they seem to be doing much smaller and slower rollouts and even some split testing. That's a fact we can observe from Teslafi data.
The majority of the FSDBeta fleet is now running a
two versions old release version
Folks still on 10.10.2 outnumber those who on 10.11 roughly 9:1.
Folks still on 10.10.2 outnumber those who have gotten 10.11.1 almost 5:1
We know Tesla has recently been no longer just tweeking existing systems with a combo of "updated perception NNs" and "update conventional planning and execution code"-- but actually increasingly creating new NNs to move more of the whole stack from conventional code to NNs.... which means results will tend to be less predictable.... and thus you want to roll out things like that slower, and to smaller test groups. Potentially even DIFFERENT version to different sections of your test fleet.
Which- hey- that's
exactly how the actual rollouts are going
In fact Green observed there's now a production and an experimental stack in the released firmwares- meaning Tesla could actually swap on the fly without a firmware update between 2 sets of code on the same car- adding even MORE flexibility to do A/B testing of newer more black-box type code.
All of which means they have ways of "expanding" the test fleet now without even needing to further add cars.
Bear in mind from a developer and company perspective the goal of the FSDBeta program could be described as:
To gather data and feedback from the minimum number of people needed to maximize progress of development.
"1 more dude in the bay area" adds nothing to the program.
While "people north of the border" does.
Why the minimum number to maximize progress? 3 reasons:
1) Limit signal to noise ratio of feedback
2) Limit staffing needed to handle that feedback
3) Limit the risk of someone ignoring the rules so badly they have an accident- which not only hurts the company in obvious and immediate ways, but might hurt them long term in broader more regulatory ones.
That minimum number will change over time.
It grew a lot in 2021.
It's growing far slower in 2022, and that is not likely to change unless you believe they're going to get regulators in the EU on-board anytime soon (and why that's super unlikely is a whole other thread).
All that said- I'm going to also give you an idea about when I think we'd be most likely to see any significant bump in NEW US invites... and this IS the one speculative (though informed by the "what we know" and "how testing generally works" stuff above).
V11.
Because up to now, they've only needed data on city streets driving.
And based on the near stop of new invites it appears they had reached their "minimum needed to maximize progress" number in the US.
But the folks in that cohort who largely do city streets driving will not be significant contributors to the data pool for highway driving. Which is what V11 is supposed to bring into the beta.
Thus they may decide they need to add a significant # of new folks- folks who the data they already HAVE thanks to them opting in for data collection via the button- tells them drive mostly highway miles.
So if anybody wants to try and stick out the safety score thing, IMHO trying to put up a fair % of highway driving is probably not a bad idea to increase your odds.
Anyway- hopefully the above finds at least few folks more interested in facts than emotion, and maybe provides them a bit deeper understanding into what might be going on beyond TESLA HATES LEGACY OWNERS conspiracies.
