Elon said the following:
"... demonstrate a demonstration"
Rule of thumb for software development 1 unit of effort for internal demonstration, 2-3 units of effort for external demonstration, 10 units of effort for production. I'd love to see as many flawed human drivers replaced by computers as possible, as soon as possible (myself included - a focused AI can exceed a general purpose system - human- in it's trained realm without question) . I just think the sensors are not there yet, and there is a depth to the last 5-10% of the task which will prove much harder than believed today.
I still think back to Gates confidently talking about speech interfaces to the PC.
Bill Gates, 7 October 1997: “What I am saying is that I’m optimistic enough to believe that
within the next decade, we will see progress to a level that for things like dealing with data in a spreadsheet or text in a word processor, or navigating the Internet, you will find the speech interface has enough accuracy that it becomes a primary way of interacting with the machine.”
Bill Gates, 16 November 1997: “I talked about some recent advances that really have me excited. Here are some that are
literally within the next few years. Speech recognition. A big, big breakthrough.”
We're getting there now with speech recognition (barely as a primary interface) after twenty years, almost a decade later than predicted. I think some of the issues are very similar. Just as we humans are able to interpret imperfect speech signals sent to us by other humans through experience, inference and visual clues - we apply our depth of experience outside of driving per see to interpret the world we see as we drive.
Take a child playing ball by the road. We all know there is a risk the ball will bounce in the road, and the child might follow it - so are likely to be more cautious (if we are observant and not distracted) . For the car to process that it's
simply a matter of teaching it to recognize the child (vs adult), the ball (in motion) and infer there is increased risk of a sudden change of direction of that 'object' by the road. This is just one of a zillion examples of how we apply 'deep understanding' of clues outside of the specialist activity of driving to our driving experience. The advantage of an AI is that it can learn from
everyone's experience, so it can learn faster - but I think the depth of learning it needs is likely grossly under-estimated at this point.
Just as 99% uptime for a computer sounds great until you realize that's more than 3.5 days downtime per year (hopefully not for an online retailer at Christmas)! 99% driving accuracy leaves a lot of 'casualties' in the time it takes to achieve the last 1% of accuracy (let alone the last 5%-10% which is the range of deep complexity I suspect will be tripping it up).
Given FSD requires constant supervision - what's the point? More importantly we have already seen some of the results of a much less capable AI 'driving' for people who chose to pass more responsibility to the AI than it was designed to have. This will be compounded by greater capability in the AI. "Dave, there might be a problem ..." is likely to be too late for the person reading Facebook, watching YouTube or replying to a text as the self-driving car decides it needs help.