I used to be in the Early Access Program with my 2019 M3 SR+. In 2019 I gave frequent feedback to Tesla on the problem you described above. I sent pictures to Tesla of the corners where the car would be alerting loudly to take over while at the same time maintaining a dangerous speed into the curve. I stressed how this made no sense and wasn't safe, and asked Tesla multiple times to at least slow the car down significantly when the car comes to a corner it can't negotiate. They listened and FSD has improved greatly.
I've since upgraded to a 2020 MY with HW 3.0 FSD. Now the car does in fact slow down and anticipate curves very well, I believe it may even read the suggested speed limit signs for tight corners, as well as the directional arrows we use here in the U.S. to indicate sharp turns. For example, I just drove from South Lake Tahoe, CA, to Fort Bragg, CA on California State Route 20, a quite curvy, hilly, undivided 55mph highway. The car performed nearly flawlessly over the entire 6.5 hour trip, slowing down smoothly and with control at tight curves, maintaining appropriate speed in the turn, and exiting back up to 55-60mph in a human-like manner. It can be a bit jerky and hesitant at times, but nevertheless I would describe the trip as nearly fully autonomous, one just has to be on the lookout for road debris and potholes. When one comes to a stop sign one has to disengage FSD and make the turn themselves before re-engaging FSD, that's about the only input I have.
I see a variety of descriptions here on TMC regarding FSD performance. I think it has to do with what FSD hardware one uses (if they have FSD at all, sometimes I question whether the poster even has FSD the way they describe their vehicle behavior), where they are driving the vehicle, how clean the cameras/sensors are, climate, road conditions, etc. I drive 99% of the time on dry, sunny, well-marked highways in Arizona, Nevada and California, where (presumably) Tesla gathers most of its data and does much of its testing. And I can say with conviction FSD on highways has improved immensely over the past two years. I'm still astounded at how well the car does on road trips, and I can't describe it to someone, one just has to experience it for themselves.
Anyway, my Tesla has for some time been able to recognize upcoming curves on undivided highways, slow down preemptively, and negotiate them in a human-like manner. The only curves it still can't do is hairpin turns and roundabouts, but it appears FSD Beta is making good progress on this front. An additional very positive safety development is that my car also slows down in areas where I, too, would naturally slow down, like with pedestrians or cyclists close to the side of the highway, anywhere where there is congestion and uncertain conditions where a reduction in speed is warranted. Again, perhaps this has to do with where I live and drive, but my point is that Tesla is making significant progress even without Dojo neural nets or whatever machine-learning technique(s) they are developing.