Probably really bad now, or doesn't react at all and also probably will collide in critical and dangerous positions. But that is also not the future. The future is how does it interact in non-dangerous positions long-term while other cars are self-driving in compliance.
Not hard to detect for the sensors, the questions is how (if) the software reacts.
He uses an artificial neural network to make the driving decisions.
You train them by showing millions of situations (images of the surroundings, information about the cars current state and information about the imminent past in this case) and the wanted reactions/solution/answer (how to drive in this case). Hotz probably has enough "normal" data, because you hold the line all the time when driving and apparently there were enough situations where the car in front slowed down and he slowed down.
From the latter case you could guess that the car would break more rapidly if the car in front stops more rapidly than in the training sessions with a human driver (you would have to test that, as you usually cannot just look at the trained neural network and see that, because they become insanely complex after a certain depth).
But probably the car never "saw" bricks falling of the truck before you or even a kid running from the left side. Those are edge cases, that are unlikely to happen, especially on a motor way. But I'd still like to know whether his software would just apply the nearest trained case and see it as noise, or notice that something is off here and alarm the driver.
Definitely a cool project, don't get me wrong. But I get kind of sad when the work of thousands of scientists who worked on the theory and the work of Tesla, Google and big car manufacturers gets dismissed by "but look at this guy" (in general, not directed at you).