Technology helps self-driving cars learn from own memories

Honestly, I think this is an important area of research for AI development in general. Every NN separates training from inference, for several good reasons, but that creates systems with a sort of short-term memory loss: they know what they learned during training, but immediately forget anything they experience during inference after. It's why, for instance, GPT-3 only knows about things up to October 2019, despite surely many prompts since then including more up-to-date information it could theoretically learn from.

Training set curation, normalization, and sanitation are extremely important, which is a major reason inference inputs aren't used to further train a network after its training is complete; but the systems would definitely see a boost in performance if we could manage to continually train them effectively somehow. A sort of online normalization and sanitation system that can also detect poor data and reject it from the training process. Perhaps as a sort of reinforcement approach? I dunno, just brainstorming; but there's definitely something here that should be explored more generally than just autonomous vehicle AI.