Wayve pushes very hard on what is called "end-to-end" approach. That means, you input sensing into the deep network and out comes the controls. No ifs and buts.

Wayve pushes very hard on what is called "end-to-end" approach. That means, you input sensing into the deep network and out comes the controls. No ifs and buts. When it came out on 2017, no one I knew believed that it can possibly work with reasonable reliability to actually put things in production. Deep learning is simply not there. You can converse with GPT-3 and find few great example to cite but within few tried you will encounter things which don't make sense. Deep learning is a great component to provide pattern recognition but there are still a lot many more pieces such as memory, abstractions, compositions etc that we have no idea how to do it. The pattern recognition can certainly emulate those pieces to certain extent but it's not the same thing. Consequently, E2E self-driving only using current state of deep learning is unlikely to produce something beyond video clip demos.