They are not within 3% or even close. They are 2,000x worse than Cruise which is still at least 8x worse than human drivers. Being generous they are 16,000x (1,600,000%) worse at unassisted, intervention-less driving than humans. To put that into perspective, if every car in the US had Tesla FSD Beta and they were all using it as a fully autonomous system without babysitting, it would average 1,750,000 deaths per day and everybody in the US would be dead in 6 months. Within a week it would kill more people than have ever died from traffic accidents in the US. We are literally talking 1 deca-Hiroshima per day of badness. The only saving grace of this whole thing is that hardly anybody is insane enough to use it unattended more than a few times so we do not see the sheer catastrophe that would occur if it was actually used as advertised.
I'm not sure flat risk per distance is the right model to use. If we consider a model like n classes of issues that all need to be handled one after another, then after most changes the incident rate will still be far above human, until it suddenly becomes much more safe. In that model, the failures come from the software failing categorically, not probabilistically.
