So let's inject a few calm thoughts here...
Over the last 3-4 years Tesla have been working on what they have now branded as "TeslaVision", which is basically the "passive" end of the AP stack that takes raw camera etc input and generates a "world-view" of the car's surroundings, both statically (where stuff is) and dynamically (where moving stuff is going). That's what Karpathy and his AI team were working on all this time. It was a huge gamble as no-one had tried anything like it before (really). And the gamble seems to have pretty much paid off. NoA with the V11 vision stack is now vastly better, phantom braking is hugely mitigated, and FSD in V11 demonstrated that the car could reliably build an accurate world view w/o lidar/radar etc (something a LOT of people, including many on these forums, had serious doubts about).
BUT, the "active" end of the V10/V11 stack sitting on top of TeslaVision was a mess: that 300,000+ line pile of C++ code doing all the active planning and driving. If you look back at the evolution of the V10.x and V11.x line, it's pretty clear that Tesla has lost control of that code base. The tell-tale signs were the long times to fix issues and the regressions as one fix caused other previously working features to break. Once code gets to this point a re-write is mandated (though if that doesnt happen the developers have job security maintaining the huge mess!).
That's what V12 is really all about. And Tesla is making another big bet, just like they did with TeslaVision .. that they can do with NNs what they could not do with traditional code. So right now, V12.x is
really V0.9 of the planner etc implemented in a brand new NN stack sitting on top of a mature TeslaVision NN stack.
So yeah,
of course it's not doing some things as well as V11. Did you expect it to be better than V11 at
everything? That's not how technology works. Its a saw-tooth .. the new approach is
bound to regress for some stuff, but the foundations have been laid for all the things that (in all probability) they would NEVER have been able to do AT ALL with the V11 stack.
V11 had reached a dead-end. This first release of V12,
even in its infancy, is already nearly as good as the BEST they could get from V11. I'm actually surprised at how good V12 is for such an early release. And that, to my mind, is a clear indication that Tesla are on the right track (which, btw, I was privately very skeptical about, given the complexity of the problem .. they have surprised me).
Those of us who have been around long enough might remember operating systems like System 7 (on Macs) and Windows 95/98 (on PC). These both represented end-of-the lines for software stacks. They were getting old, and new features were hard or impossible to add to the aging and byzantine code bases. Both Apple and Microsoft were forced to re-write their respective OSes from scratch: OS X (macOS) for Apple and Windows NT (now Windows 11) for Microsoft. The early releases of those operating systems were
serious feature regressions compared to the "last gasp" of their predecessors (slower, lacking in features, compatibility etc). BUT they laid the strong foundations on which Apple/MS could continue to innovate, and their newer versions are
orders of magnitude better than System 7 or Windows 9x could ever have been. That's where we are now with V12 .. the
very first release of the next generation, with the potential to grow WAY beyond the older V11 work-horse.
imho, it's going to be a very interesting year
