On the software improvements, for me a threshold on usefulness is the removal of obvious flaws to a low enough level to gain human trust. Ie when we think the failing common sense answers are less than 10% can we reduce that to <1% and then <0.1% with software improvements alone (or perhaps software +the next hardware upgrade that Nvidia is currently producing) . My analogy for the trust question is self driving cars where YouTube video makers worry that without near crashes the have nothing to make a video that people will watch. I think this was v13 software and v3or 4 hardware for tesla. Anecdotally I see people retrying the latest models after a year or so since ChatGPT3.5 with more success, so take up for anything useful will be fast as the error rate decreases pass the trust threshold- but do people here think software improvements are enough to pass the trust threshold?
I think the question of ai controlled humanoid robots vs humans and horses vs cars is interesting. RethinkX on this subject is kinda scary. Humanoid robots amortized over 3 years and say 16-20 hours a day quickly go below minimum wage in America so the question of whether the economy has enough tasks for such a labor revolution that starts out not very general, but becomes more general use over time. So I encourage your thoughts on this topic, as I don’t know what framework to use to think about it