Do you work at Tesla or somehow have firsthand knowledge of their engineering system? I don’t, but I am a Principal software architect at Microsoft who specializes in both modern software engineering + delivery practices and machine learning (particularly on-device ML and end-to-end feedback loop systems like what Tesla uses). I only bring this up because as long as we’re both speculating, I’m going to continue using my experience to inform my speculation.
Parts of your explanation are likely accurate, but other parts are almost certainly not. I think it is very unlikely that their data collection and retraining systems are automatically spitting out new outputs and automatically pushing them into the next builds from master. I won’t say it’s impossible, but we have no information pointing to that, and I would put my money on the NN pieces being trained, experimented with, and validated in an independent system and only when something is available which scores higher on their metrics is it submitted (e.g. via Pull Request from a topic branch) into master. There is no requirement that this happen every week, and I’m quite skeptical that it does in practice.
