"neural nets helped speech recognition make big gain"
I've not noticed much change in voice recognition QUALITY in last 15 years.
" but all it had were 20-year-old techniques."
Actually thirty year old techniques. It's still all there is, except faster (more powerful CPU / GPU adds zero AI, only speed) and with bigger databases.
" to harness neural networks (now rechristened as the much sexier "deep learning") on narrow demo-friendly challenges such as image recognition and text processing,"
Perhaps because they are nothing like biological neural networks. It's a data flow design based algorithm approach (invented in early 1980s) well suited to multiple cores (first popular on Transputers) and distributed systems.
Though curiously I can't find a grammar checker or spelling checker much better that I had on CP/M and DOS 1987 to 1991.
Google translate is a LESS advanced idea than translation in 1990s as at its "heart" it uses a Rosetta stone brute force approach. They started with EU texts. Not by improved parsing and grammar models.
Industrialisation and automation dates back to the end of the 18th C. with powered programmable looms. CPUs and programs have replaced 1960s relays and ladder logic. It's been incremental over the last 50 years in electronics assembly.
Most AI is hype, marketing and Humpty Dumpty "Words mean what I want them to mean". It's brittle with the unexpected as it's ultimately human curated rules, data and complex human written programs.
We have no idea why in some respects a crow is smarter than a chimp. Why Corvids, budgies, elephants, chimps (only sign), dolphins, starlings and parrots can learn vocabulary as symbols for real things and can't do language.
We have no satisfactory definition of exactly what Intelligence is, which seriously hampers program design. We are no closer to General machine intelligence than 60 years ago. All progress in AI has been by innovating a machine solution based on brute force and bigger data sets (to reduce the run time brute force needed, e.g. previous chess matches). Redefinition of marketing terms from database, Expert System, AI, Neural Network, deep learning.
Yes machines replace human jobs, not new and arguably you should look at 1837 to 2017, not just since 1997. It's slowed a lot.