Remember, these cars are running local LLMs, not frontier models. The issue with self-driving cars has been the edge cases. The 0.0001% of situations where the models did not have sufficient training data. This is compounded by the hardware limitations. Onboard RAM in a typical Tesla on the road is 16GB (+16GB for the backup computer). This has to run the existing onboard OS and other operations plus the LLM. These two factors combined means that the cars are currently incapable of negotiating the 0.01% cases, let alone the 0.0001% cases. And this is compounded by the fact that LLMs cannot currently update their weights in real-time, like humans. It takes months to train a new model. Special small models can be very tricky, especially around safety and mission critical applications like FSD.

All that said, current data shows that FSD is already better than human drivers on average. See the recent regulatory decisions by the Dutch and Danish road safety authorities. So we've already crossed the rubicon. All improvements now are icing on the cake. My prediction is that local LLMs will get much better, very fast. How that's operationalised with Tesla (or other) data is yet to be seen. They have at least three new ASCIs/SoCs in the roadmap for improved LLM efficiency and with a lot more RAM. Plus they just announced new technologies allowing the local LLMs to learn from driver intervention and behaviour. Some form of vectorised RAG, which could mitigate a lot of the limitations around real-time learning.

I am very optimistic for the future of self driving. I own a Tesla with FSD now, and it's incredible. It makes mistakes, but fewer than I do, and so far has saved my butt (and my wife's) several times from obstacles and emergencies we would not have seen. The car has undeniably made us safer.