In practical terms it means that Tesla is no longer hand-crafting the heuristics to control the car. Instead, they're using a neural network, which can be trained to figure out control heuristics by, well, instinct. Instead of a bunch of software engineers working long and hard for years on ways to create clever heuristics that cover every conceivable scenario, the system can be told "Do it like this" and then be shown zillions of scenarios of "proper driving". The system just figures out how to respond to scenarios.
The outcome that we're all hoping for is that FSD will break free of its log jam of poor decision-making. The hand-built heuristics had reached a peak and were no longer providing much in the way of improvements. We want neural networks to inject some machine learning magic that turns FSD from a frightened teen driver into a competent chauffeur.
The end-to-end term is used to say that all of the steps in the decision making process are now neural networks. The current solution is many steps of neural networks, but the last step is not. The last step is C++ code hand-written by engineers. Once Tesla has replaced that last step with a neural network, it'll be entirely neural networks; all neural networks from start-to-finish - end-to-end.
There is also a use of that term which means that the entire decision making process is covered by one big neural network. Folks argue about which one is really end-to-end and which one Tesla is pursuing. I think the consensus is that, for now, Tesla is still going with a bunch of steps, but all of them are neural networks.
