Many large companies have research departments that do experimental work that'll never get to the product. This raises prestige, increases visibility and helps hire smart people.

Things like Waymo's EMMA is an example of this. Will the production cars use LVLM's somewhere? Sure, probably a great idea for things like sign recognition. Will they use a single end-to-end model for all driving, like EMMA? Hell no.

Driving vehicles with people on board requires an extremely reliable software, and LLMs are nowhere close to this. Instead, it'd be usual layered software - LLM, traditional AI models, and tons of hardcoded logic.

(This all only applies to places where failure is critical. All that logic is expensive to write, so if there is no loss of life involved, people will do all sorts of crazy things, including end-to-end models)