by Outcast_Searcher » Sat 20 Apr 2019, 17:34:33
Getting back to the thread subject, the interesting place they've made really BIG progress in recent years re AI, is in playing games. i.e. very complex but well defined games.
AlphaGo and AlphaZero, redefining the way an AI is built, and taking the underlying theory of the go and chess games beyond what humans were coming up with. And then, extending that to AlphaStar, to dominate the highly popular Starcraft II game, where there is a well defined game space and rules, but no longer a finite number of positions or "board" states.
You might argue that this isn't "thinking' at all, but if you know a fair amount about the theory of such games, and watch the AI, trained up over a couple-few months of intensive iterative self-teaching via experience, it's DAMN impressive and spooky to watch them obliterate the best humans at such games.
This is a COMPLETELY different approach than the brute-force search method that worked for chess over several decades mostly by throwing enough hardware at it -- but was failing MISERABLY at a game like Go, where it's MUCH harder to accurately define aspects of the game with numbers (at least by top human players) and where the game space is gigantic compared to chess.
I'm not sure how quickly that sort of effort finds its way into real-world robotics/AI, but it seems entirely logical for applications with a well defined "board" (sphere of operations), and goals. I think that a whole class of jobs like packaging, or basic food preparation, etc. are likely to come under direct assault by such algorithms within the next decade.
Now, Tesla fanbois are trying to use these programs as examples of why Musk is right that true level 4 or even 5 full self driving AI will be reality in 2019, re Elon Musk's recent claims. (Even though his self driving claims/predictions for 2016 - 18 were total nonsense and failed completely everywhere but in Elon's excuses).
I call total BS on that, since the world and all its roads, and everything unexpected that can happen on those roads is NOT a well defined or remotely simple thing to understand, much less "solve". Even the known and expected issues like weather, visibility, idiotic drivers, idiotic pedestrians make Go or even something like Star Craft II look like learning to count to 7.
Now, give such processes a decade or three to mature and be studied and advanced for FSD, who knows? But of course, that's the kind of time frame being predicted by many of the experts trying to make it work already. (Those who have credibility, unlike Musk et al).
Given the track record of the perma-doomer blogs, I wouldn't bet a fast crash doomer's money on their predictions.