The decline of computers as a general-purpose technology (2021)

I largely don't blame the specialized processors, they exist to fill a need. The transition to mobile/tablet has been very much a producer-consumer relationship. In the old days, normal users would one day "View source" and dabble in HTML and making their own web pages and journey into making their own programs. There's no parallel to this on mobile today. Perhaps the next generation that grew up on Roblox and using yet to successful end-user visual development tools will.

Is there a web-app like MS Access or Visual Basic today, maybe Airtable, others?

It's way too premature to say we are turning into fragmentation cycle regarding bespoke hardware. GPUs are still universal computers, they just use a different model with different performance characteristics (and applications, which stem from these performance characteristics and also inertia/maturity of tooling) than CPUs. ASICs and specialized processors have been a thing for a long time and their application towards some "new" things like crypto, hft, and a handful of cloud optimizations is hardly a new trend when seen in the larger context of their being used all the time for now-boring, once also-new technologies like networking hardware and embedded applications.

I'd argue that the last three decades have actually had massive shifts towards more generalized and ubiquitous computing that may continue even further. First, with the introduction of PCs the computing landscape shifted away from bespoke hardware and software to more standardized plug-and-play hardware and software. With the internet, everybody got a middleware called a "web browser" that standardized UI development on a single computing model. With mobile computing, we got a more restricted model of userspace that standardized what applications could do and even how (app review, software signing) you could do it. Cloud did the same things to various extents to server software. Except for at the height of Windows' PC dominance, it's never been easier to write software usable by such a large portion of the market, and the absolute number of devices and users that applies to is probably an order of magnitude more than existed during peak-Windows.

Everybody in tech knows single-core cpu performance gains are slowing down and that the remaining incremental improvements will eventually end too, so what comes next is on peoples' minds. IMO this article is jumping the gun though - it's still way too big an undertaking to use specialized computers (and I don't count GPUs as these) for all but the largest scale or most performance sensitive use cases, as it's always been.

But surely she had the tools to do that in excel itself, no? This is something I see a lot in software, where people devalue and trivialize the skills and abilities needed to make software simply because they know something is possible to do in software and because, as a software worker, they most likely had a natural predilection towards the kind of thought processes that led them toward pursuing software work.

I think it's fallacious to think everybody has the natural disposition and curiosity to want to write software. Writing has been around forever, and most people don't want to do it beyond purely functional communication, and even those who do want to go beyond that are mostly terrible at it. That lady could have automated her job if she really wanted to and was motivated enough to do it; she simply preferred to be a ditch digger than to design a ditch digging machine.

It doesn't help that "designing ditch digging machines" is in such high demand that most companies are really only able to hire mediocre people to fill those roles. Leonardo da Vinci is not working on sox compliance at a fucking bank.

There's a good chance that the payoff for that bank automating "take a set of dates in an Excel file, search each one individually into an internal search tool, then Ctrl+f the result report and verify that every code listed in the original document exists in the report" is not even there, because the first attempt will have insufficient monitoring or verifications to catch silent failures and introduce an expensive issue caught months after the fact that leads to a huge remediation scramble and requires more development than initially expected to address, not to mention the constant need for maintenance in updating the software and making sure it continues to run (which also involves a rotating cast of programmers joining, needing to learn the software, and leaving). Leonardo might do it right the first time, but he's not doing it for cheap and not going to make it his life's work. Joe Schmoe can do it cheaper but it's going to take a while, have an ongoing maintenace burden, and probably some mistakes made along the way. That lady can do it for even cheaper than Joe Schmoe can without many headaches.

Back in 2009, I visited Tokyo for work. What I learned from the folks in my local office is the deployment software my company made tooling for was not very popular.

The reasons were mainly that most of that software was made by Western (English speaking, specifically) countries and they did not trust it. They were not exactly wrong - software tends to work best on English systems, unless it was developed by a local company. I heard of some really embarrassing Korean translation issues that even as a person just with Google as my Korean language skill, I could validate the translation issue. Like Korean 101 kinda mistakes.

So Japanese companies were using vastly inferior deployment software, which would basically RDP into an interactive session and replay inputs, because that software worked fine in Japanese.

...Or also very common was sneakernet deployment with DVD's and thumbdrives, making the lowest workers do the running around.

*I do not know the current state of software deployment in the Japanese market.

That is what I had in mind as I wrote it. Sad that many macOS apps now are built with Electron, or we simply use web browsers to access public or internal apps.