This is true but there is a reason I became a lot less optimistic about the timeline for full autonomy sometime around the middle of last year. Basically, it's due to the flattening of the improvement curve, a curve that was previously steepening. Trends matter. Previously, the improvement curve was accelerating, and I expected that to continue. That was my failing. The reason it did not, IMO, is we have entered the "march of 9's" and I think it's under-appreciated how significant the "march of 9's" really is. Even Musk seems to know this judging by the fact that most of his specific near-term statements about FSD timelines exclude regulatory approval and actual autonomous deployment, choosing to focus instead on "feature complete" or other metrics that are before the "march of 9's" is fully complete.
Because I think the bulk of the value of FSD does not happen until it is fully autonomous and, even then, it will take a while for the general public to accept its superior safety, I think most of the value will be recognized and capitalized on in a more gradual manner, over a longer period of time, not in a singular "aha" moment. FSD is already contributing to auto margins and that will gradually accelerate with increasing levels of adoption, increasing levels of revenue recognition and, eventually, greatly increasing prices and real-world, profitable applications. But it will be a gradual process, overall, and there is no one defining moment at which it is perfected. That is what the "march of 9's" is. A long, gradual process.
I do have some hope that Dojo will greatly accelerate the "march of 9's", hopefully starting later this year, but that doesn't mean it will be a fast thing, simply that Tesla has increasingly powerful tools for a very difficult job. That said, people who think full autonomy will take until 2030 or beyond are just out in left field without sufficient imagination to see that the foundations have been laid for full success, well before this decade lapses.
I believe your sentiment regarding the flattening of the improvement curve of FSD(beta) capability is shared by many Tesla investors/followers.
I too have adopted a more pessimistic view in the sense that autonomy is "less than five years away" instead of "less than two years away" (when Elon would spread hype with bold statements on Lex Fridman/AI Day #1/Autonomy Day).
That said, I'm still more hopeful than some given the following.
When talking the flattening of the improvement curve, Elon always points out that improvements follow an S-curve. This is true for vehicle manufacturing ramps, but the same is true for software improvement. As Elon mentioned (for example in his last interview on the Lex Fridman podcast if I'm not mistaken), sometimes a lot of work is necessary to take one next step, but that step builds a foundation for easier improvement after that step. Therefore sometimes we see little to no improvement (or even regression) for some time, but foundations are built for faster improvement and a higher performance ceiling at a later date.
Some examples of this in FSDbeta development are the code rewrites to for example (1) move from processed image input to raw photon count, (2) move to single stack for FSDbeta v11, (3) move from single image recognition per camera to fusing camera inputs, fusing by Neural Networks instead of heuristics based, ...
Right now the Tesla Autopilot team has just rewritten the entire codebase to transition to the single stack, and we saw very little improvement in the recent months (or slower than most had hoped). As predicted by many, including Elon, the freshly built single stack to rule them all would be a step backwards on highways from the current Navigate on Autopilot capability, and most likely Tesla saw it was a huge step backwards and did not want to release v11 in this state.
The many v11 releases with employees only are IMO a race to feature parity with FSD beta v10/Navigate-on-Autopilot.
However, once v11 will be released to all curent FSDbeta participants (400k in the USA), the software can again improve at an increased rate.
Not that v11 will magically provide perfect autonomy, but that its performance ceiling will be much higher than the current v10 (which splits up driving tasks depending on location).
Most likely the v11 - which among other things will make use of the Occupancy Network (currently active for city streets in FSDbeta v10) - will be the foundation for park assist features based on Occupancy Network to run on cars without ultrasonic sensors. I'm guessing the Occupancy Network is currently being improved upon before v11 release as to make it reliable enough to provide park assist/autopark and the like. If not this would result in possible liability since park assist features aren't just beta software with little legal consequences.
To sum up, I believe the past few months have been the end of an S-curve, but also mark the beginning of a new S-curve in software improvement that will find support in the single stack. I'm very curious to see where FSDbeta will be at in 6 to 12 months.
(Final note: I'm less optimistic on Dojo in the short term, since Elon hopes for it to be 'only' as good as the GPU cluster this year, and better than the GPU cluster starting 2024. Since Dojo is just another method for video training, it is no magic bullet. Yes, Tesla will be able to train more NN's with both the GPU cluster AND Dojo by end of 2023 but that's only a 2x improvement. I'm more excited for when Dojo is 5x/10x better than the GPU cluster and Tesla can train ridiculously fast compared to nowadays.)
