We can tell the "controls" part is much smoother in V12 - the part that was heuristics previously. That's a great sign. Right now disengagements in V12 are higher than V11, but I'd guess that will match V11 the first half of this year. Look at how long its been since Elon's V12 demo till now - 5 to 6 6 months. Good progress, but not like it's mindblowing. So I'd expect the 2nd half of this year for V12 iterations to show disengagements at maybe 1/2 to 1/4 of V11.
Combined with the increased smoothness is probably enough to have a L2 product that is quality enough many people would pay some money for. The equivalent of autopilot for city streets.
But Tesla may have to lower subscription prices to engage more people - I would think this could increase overall FSD revenue and think it's an obivous move to boost margins a bit. Maybe they are waiting until it's good enough to do this?
The thing is, great L2 will boost margins a bit, but doesn't make life altering changes to the company valuation. That only happens if people think robotaxis are imminent. Realistically, there is a long road between great L2 and robotaxis. The disengagement rate has to be really, really low to be software capable for robotaxis. Like 1 critical disengagement every 10,000 - 100,000 miles. Not every 100-500 miles like currently.
So I'd rate there almost 0 chance Tesla gets close to that level (which the market would definitely respond to).
The real question is: can Tesla get to a point that the pace of change is so rapid that the market prices in future expectations years early? This is possible, but the rate of change would have to be much faster than it is currently. Part of that depends on how much compute do they have coming online. Last we heard it seemed like they 2x their compute (only). Way below the 10x-100x they are planning for. I was disappointed Elon did not respond to the question about compute limits on the earnings call. This insight is needed for any institutional investor to get more bullish.