|
AI doesn't scale well. Problems get worst as you make your model bigger and more generalized. To make things worst, data, model architecture, precision, hardware, affect your model performance in ways that are hard or impossible to anticipate. If you watch Tesla's AI presentation, https://www.youtube.com/watch?v=HUP6Z5voiS8, you will notice that they have multiple AI's stacked on each other, which IMO is a step back from truly e2e multimodal AI system. So even with their custom fancy hardware, multimodal is too hard. I wonder, wouldn't it be better to use geo fencing (using H3), and have the car download the model depending on the zone where it is driving? And optimize multiple models based on "driver engagements"? This could fix the problem of zones where there are particularities in the driving, road, or human activities, and allow for model optimization to happen on a smaller vector space than the whole world. For example, why not have a model for US highways, LA, New Deli, UK, so on. Tesla also knows where the cars are, and control their expansion plans worldwide, which could inform model prioritization roadmap. In my mind, it will be easier to test, debug, label, optimize, and guarantee quality to users, that at the end of the day, without knowing exact statistics, I am dare to say spend more than 70% of the time driving around the same county/city/area/town? |