The key issue is that ml/dl is pure statistics - there is no intelligence or learning or conceptual awareness of space-time, so that technology can never do so many things people try to do with it.
Some ml/dl tools (think gpt3) seem to be able to answer questions that we previously thought you needed gintelligence for. I think the line between "pure statistics" and "intelligence" are much more blurry than they used to be, and might go away entirely.
Actually, I think the industry is finally realizing that there is no intelligence there, especially with gpt3 which can figure out with great precision what statistically comes next, but there is zero understanding of the space-time conceptual meaning for gpt3 in that answer - its not designed to do anything but figure out statistically what is most likely to come next.
> Another key piece of evidence, the failure of all FSD attempts trying to use ml/dl thinking its more than just statistics.
I would almost put FSD as a good example here. Yes, some attempts here are very naive and try to use ML as a magic black box tries to covers a long stretch of the system from vision to turning the steering wheel. However the best performers just utilize ML for small well-defined parts of the system and in a "statistics on steroids" way with most of the other parts utilizing much more traditional methods.
Some of the things GPT3 can achieve are very impressive, but once you work you work with it a little bit, you definitely feel that it just regurgitates the masses of text it has been trained on and tries to piece it together in the most cohesive ways. And for production usage, it (and similar models) have huge problems with halucination where it will confidently spit out "facts" that are just plain wrong.
That’s what many people do as well, I feel. Not joking.
That aside, copilot works well as a proactive search tool for idioms, templates, and boilerplate. It can also do things like build the invocation of a CLI tool, with arguments, from a comment line. To me, if nothing more, is the next step in the use of reference. I started with man, .hlp files and books… then progressively replaced them with Google because it was faster even if more noisy… now I use copilot first.
