Large language models are totally unsuitable to autonomous vehicles. No one is going to write code to describe a dynamic driving situation as static text which could be used as an LLM prompt.
Some of the underlying core technology could theoretically be used in that way by taking input from multiple videocameras and other sensors. But that approach is a dead end for consumer product development until someone solves the testing and explainability problem. Current limited AI systems are too unpredictable for safety critical applications. Just because AI software seems to work in testing doesn't mean you can have any confidence that it won't do something wildly dangerous in a novel real world situation. Vehicle manufacturers and their insurers will not take on the high potential liability even if autonomous systems are on average safer than human drivers.
Tesla uses a small language model for lane selection at intersections, they already shipped that to customers.
Natural language is useful for systems with discrete elements and complex rules, which seems like a good fit for parsing intersections, construction zones, things like that.
I also disagree that the explainability problem needs to be solved before self driving cars can be implemented. The effects of drugs are inherently non-deterministic, but the FDA still lets companies develop drugs and ship them to customers. You test self driving cars the same way we test drugs: by collecting data and calculating the probability of harm.
Some of the underlying core technology could theoretically be used in that way by taking input from multiple videocameras and other sensors. But that approach is a dead end for consumer product development until someone solves the testing and explainability problem. Current limited AI systems are too unpredictable for safety critical applications. Just because AI software seems to work in testing doesn't mean you can have any confidence that it won't do something wildly dangerous in a novel real world situation. Vehicle manufacturers and their insurers will not take on the high potential liability even if autonomous systems are on average safer than human drivers.