Consider this more a wild speculative rambling than actual thorough evaluation, so take that grain of salt (I do work in AI/vision research):
The probability that they can safely pull this off is very low.
Right now progress in vision-based ML is incredible. If you look at AK's twitter feed [1], you'll be amazed almost every day. But I work in planning, and for planning detecting / segmenting objects, even recreating 3d models is not enough, you need a full contextual model of the world. Right now, I have 0 clue how this could be done, I don't think anyone else does either.
Here's an example of what I mean:
you see an image, and can say 'there's a person there' - that's detection.
you see an image, and can sort of imagine the person's shape from another viewpoint - that's 3d reconstruction / completion.
you see an image, and can imagine the person moving, obeying social and physical laws, waiting to cross a street, look at a car and decide to wait. You know trains will move on rails, trees will stay still unless cut, etc. You're mixing physics engine, statistical inference, heuristics in your head, and this gives you a full, contextual understanding of the world.
I think to get there, one approach is to focus on prediction ability: create a model that learns to 'see', then 'predict' and train both at once (see for example World Models [2]). This prediction task alone was also enough for GPT-3 to seemingly learn really good models of the world. Then the hope is, if you can predict the future really, really well, somewhere in you must be a really perception / simulation / contextual understanding model. In public research we are so incredibly far from being there, prediction ([5], [7]) and prediction-based planning works (eg [3], [4], [6]) are still in the proof-of-concept stage - we're nowhere near predicting high-res images or 3d models of the world with any accuracy. Still, I do see some chance of a GPT-scale endeavor by a well-funded group like tesla leading to some big leap forward. But I'm not holding my breath. To make things bleaker, some think prediction is not enough, and you actually need embodiment (skin-in-the-game), to truly develop good contextual models (I used to be in this camp, GPT made me reconsider). This is even further away, adding hardware, safe exploration and co. is just not something we're in any way close to. Another common thought is that there's a limit to current NN architectures, and that a drastic step is needed to get to the next level, though no one knows exactly what (moving away from feed-forward architectures and backprop, towards something like spike-time-dependent plasticity? higher level meta network coordinators, that decide when to connect trained models to each other's inputs and outputs, when to start and stop modifying their weights, etc, continually?)
You could argue that it might be possible to skip this entirely, and just focus on training an algorithm to be good at driving from start to finish, but there I'd say that we are even further from having the approaches to reach that goal. I'll leave that topic for someone else.
And if you say you don't need this, then you better have some magical over-engineered planning method that can perfectly account for all possible errors / missing info in your perception pipeline. As far as I'm aware, Tesla has no edge there. I'd expect that's the approach other big manufacturers were throwing millions at for all these years, before ML suddenly became more than just edge detection / segmentation.
[1]: https://twitter.com/ak92501?ref_src=twsrc%5Egoogle%7Ctwcamp%...
[2]: https://worldmodels.github.io/
[3]: https://www.youtube.com/watch?v=w6TLbv_54GY
[4]: https://arxiv.org/pdf/2012.04406.pdf
[5]: https://arxiv.org/pdf/1808.06601.pdf
[6]: https://arxiv.org/pdf/1812.00568.pdf
[7]: https://www.youtube.com/watch?v=3UZzu4UQLcI
Note:
I do think they could pull FSD on motorways in a realistic amount of time, though, because there's a lot less variation there. You could even add a little bit of external control, like road-side accident detection cameras, full walls to avoid animals / objects going in and out, etc.