Look at what this software put after you sent your last...thank you. What I clipped above is what didn't get displayed. All I saw was two YouTube vids with one text line in between.
So your validation set is the set of disengagements. It sounds like you are agreeing with me or is that too simplistic. Let's take the case of disengagements. What is the feedback loop there. More training data or more if-then-else coding?
Cutting to the chase. Is it, in your opinion, feasible that auto steer, NoA, autopark, auto retrieve driver can be feature complete (the software definition) yet still not be deemed ready for SAE level 5 designation? Seems to me, for instance, recognizing a person walking away from you but not recognizing a jogger coming towards you is a training issue not an if-then-else issue.
On the vids. I was also one of the leads in the search development at Monster. Much of that video was in the domain of Lucene and Mahout. Neither I would consider machine learning.
I would say probably a combination of training data and if-then-else coding as well as data labeling. Given Musks statement about EAP I would guess they are doing a lot of parking lot data labeling, which is then fed in to systems for training models. Somewhere in there there is most likely some if-then-else coding in regards of how to handle varying degrees of certainty and such.
When it comes to whether or not I think FSD will happen I think it important to offer my perception of what that is:
To me FSD is basically what NoA does now, but on city streets. Given the complexity of city streets compared to highways there will be some additional complexity to how it handles uncertainty (stop immediately and require human intervention or veer away from a uncertain path/obstacle but continuing driving to prevent getting rear ended). The car will effectively be able to drive 90% of the time, but will occasionally say "oh *sugar* I don't know what is going on, take over for me", and do so in a safe manner. I expect you could go multiple days to and from work, the store and your kids school with FC FSD without ever having to take over. But then occasionally something will happen that will require you to do something. These moments when people take over will teach the system a lot, which will reduce the frequency of these events.
This system will not legally be able to drive itself with no passenger because regulators are not engineers, and they will want to see millions of miles in the real world validated by proof that drivers don't need to disengage AP. Also so that an empty car doesn't get confused and cause traffic jams by getting stuck in front of a pothole or strange street art.
When it comes to whether I think that will happen by the end of the year, beats me. I think the current AP version that everyone is using is completely irrelevant to that question, and the only people who can really speak to that are either Tesla employees working in that department, or the people at Autonomy day who trialed the development version. Given Tasha Keeney of Ark Invests opinion I would say my definition of FC FSD could well happen by the end of the year, but I don't base that off of anything other than the fact that people who seem to know what they are talking about who were given first hand experience seem to be saying. I don't think anyone with any amount of industry knowledge that has not had first hand experience with the development version of autopilot can honestly claim to have any more knowledge than anyone else on the forum. Even given how technical Autonomy day was they were very vague as to how the system works and how well it is performing on corner cases that would be of concern. And for good reason, thats the secret sauce.
So short answer, since the big money with in depth knowledge on the subject and first hand experience with not only Tesla but many other self driving plays seems confident in FC soon and Taxis in 1-3 years, I think the wise answer is to believe them. Nobody outside of either of those circles can fairly claim a superior perspective even if there were the foremost expert in NN and Machine Learning at MIT, because the most important parts of knowing one companies specific progress is for the most part only knowable with first hand experience.
How does SAE define feature complete? I differentiate feature complete from training complete.
The other thing is, I believe Elon Musk, whether one loves him or hates him is far closer to the state of FSD at Tesla than any of us. So the choices are one of two things, he is either with full knowledge and forethought bold face lying to owners, investors, investor advisory firms, and the street at large -- or -- he is making a factually truthful statement.
Basically I agree. I think the Robo Taxi timeline is probably wishful thinking because thats a regulatory issue and not a engineering one, but I believe Musk sincerely believe FC by end of year and has a logical reason to believe that based on evidence rather than random numbers.
