@thomasfuchs

Back then:

Computers were getting faster at such a rate that a two-year-old computer felt like it needed replacement.
Core software was gaining features that were actually useful as the underlying computing power increased.
The number of people who didn't have a computer was huge and so the total number of computers being sold was increasing each year.

All of these meant that selling computers or software was basically a money-printing machine. People wanted the newer version and the number of people buying each version grew.

Sadly, the entire industry became addicted to this revenue-growth-without-trying model. And that started to hit a brick wall when a couple of things happened:

First, software became 'good enough' for a bunch of things. I think Office '97 was good enough for probably 90% of home users. The earlier versions were missing features because computers weren't fast enough to be able to handle them. Every new release came with features that people really wanted but couldn't have within the constraints of what was possible to implement. By the late '90s, the hardware was no longer the primary constraint.

Second, hardware stopped getting rapidly faster. My 2013 laptop is slower than my 2023 one. But it's still usable. It runs more or less the same software (the old one is x86-64, the new one AArch64, but I run the same programs). The old one was a quad-core Haswell CPU running at up to 3.5 GHz. The new one has 12 cores (8 fast, 4 slow) and is faster per core. Most workloads have CPU usage sitting at under 10%, so even a 10x speedup wouldn't be noticeable for those, and it's probably closer to 5x for most things that are CPU-limited. Building LLVM on the newer machine is great, it was a bit slow on the older one (but still okay)

In 2003, a 1993 computer would have been painfully slow. A decent 1993 computer would be a 60-66MHz 486 or Pentium. A 2003 machine would have been over 1 GHz and with more instructions per clock, so a factor of 25-50x faster.

Even the gap between the 1993 machine and 1996 expected performance would have been big. The computer I got in 1996 was a 200 MHz Pentium clone with 32 MiB of RAM. A typical 486DX2-66 came with 8 MiB, so the delta from 1993 to 1996 was close to the delta between my 2013 and 2023 machines.

This combination meant that hardware and software refresh slowed down. At the same time, markets became saturated. Bill Gates' goal was 'a computer on every desk'. For most rich countries, that was basically achieved by the early 2000s. Until that point, companies were seeing rapid growth simply by more of the people who could afford computers deciding to buy one. After that, not so much. Some growth continued as the global middle class got bigger, but that grew far more slowly.

EDIT: That 1993 computer, if it ran Office, would have run Word 2.0 or 6.0. 6.0 was the first version to add spell checking as you typed (prior to that, you had to run the spell checker manually and inspect each thing it flagged then explicitly add, ignore or fix it. 6.0 really didn't like to run with less than 8 MiB of RAM (it was laggy and slow if you tried). A 66 MHz 486 and 8 MiB of RAM was fine as long as it was the only program you were running. Start up Paint Shop Pro and open a large image and you're swapping again. My 1996 machine could run Word and an image editor in NT 4 (memory protection, yay!) without swapping. That got a lot better once I bumped the RAM to 64 MiB. By 2003, the idea of having to close a program when I opened another to make sure I had enough RAM had basically gone, I closed programs when I was done with them.

I actually like to write in Word 95. It’s kind of perfect. 😄

ALT

@david_chisnall @thomasfuchs

To add my 0.5 cent to this, I think the model applied by “money people” of the software companies (and sometime sadly inside opensource communities) is wrongly the same as the one in let’s say “spoon production companies”.
The difference is to sell the 2nd, 3rd and 20th spoon you still have (more or less) the same cost of production!
In software production, once it’s made you have infinite copies to sell… so we have to accept it’s done/finished and just maintain…

@david_chisnall @thomasfuchs Typical 1995 setup with somewhat more limited budget (e.g. school): MS Works on a 4 MiB RAM 486, where Wordperfect 6 under Windows 3.11 would be sluggish. Not to mention a ~15 year old microcomputer, as in this Crow Road epsiode from 1996. https://www.youtube.com/watch?v=WjQDSVdW56I&t=2614s

@david_chisnall @thomasfuchs Sounds similar to what Benjamin Graham described in ‘The Intelligent Investor’ that happened to aeroplane manufacturers: jet planes had become fast enough that pushing for speed no longer paid off. A flight of eight hours vs one of five still amounts to a day lost to travelling. Few people were going to pay a premium for the quicker flight, so no airline was going to buy faster aeroplanes. From then on, innovation switched to passengers/litre (either fuel or cabin space). Trying to find the minimum amount of room that people will tolerate.

I don’t know what the equivalent of passengers/litre would be for computers. Trying to find the maximum amount of adverts and privacy invasion that people will tolerate?

@david_chisnall @thomasfuchs OMG where do I even begin? Memory protection as a 1996 "yay" when Sun shipped it on a desktop in 1982?

But mostly: a SPARCstation 2 or an Indy shipped with 16 to 32 MB as a baseline in 1990 to 1993. Not because these vendors were frugal (Solaris and Motif gobbled memory too) but because they sold a complete engineered system, not an artificially cheap box you'd finish upgrading from a catalog.

Refresh cycles logically followed depreciation schedules and 3-5 year leases, not an Office release buy cycle driven by a product manager's new Porsche and a mortgage on a mountaintop mansion (true story). The OS came bundled (Solaris, IRIX, AIX, HP-UX) and was never a separate revenue line, so there was no feature treadmill to drive sales.

The Wintel model was predatory and co-evolved, "Andy giveth and Bill taketh away," with cheap third-party SIMMs as the "medicine" market. The 1996 DRAM crash (Korean capacity overshooting the 16 Mb transition, prices falling from roughly $30/MB to under $8 in about a year) is what turned "just add RAM" into an upgrade addiction.

That means the 32 to 64 MB upgrade in 1996 was enabled by a glut. It's not a timeless pattern. The question is still who captures the margin on memory, not whether it is needed.

@david_chisnall @thomasfuchs I normally mark XP, Mac OS X, "usable" Linux as another milestone. IMHO those were the first "workstation-worthy" OSes that you could get at home. They also required more resources (prior NT was not really used at home due to expensive RAM requirements?).

IMHO, after "Internet at home", the other major innovation was "Internet all the time in our pockets", and increasingly people believe that was a mistake...

@coder @thomasfuchs @david_chisnall

By my recollection, NT4 had very little support for the kinds of hardware people used at home, even if you could make it work with a relatively inexpensive RAM upgrade. 2000 was much better in that regard, but it took until XP for common consumer hardware to be well-supported.

@thomasfuchs @david_chisnall I know I'm likely looking past with rose-tinted glasses and claiming a golden age, but I don't think we're really making that much "impactful" progress.

I have computers from 2016-2017 that are still quite good! (Except with bloated Electron apps.)

@ArtHarg @david_chisnall @thomasfuchs Performance:watt, so very similar to your planes.

@woe2you @david_chisnall @thomasfuchs I don’t think that that is the equivalent metric. Extra performance is no longer a real benefit and watts are not a limiting factor (for the average end user, at least). So extra performance or less watts are not going to sell computers by themselves.

@ArtHarg @woe2you @thomasfuchs

Yes, it's more about diminishing returns and closely related to Amdahl's Law.

If I fly to New York, it takes me an hour or so to get to the airport, I then want to allow a couple of hours slack, then it takes a while to go through immigration, customs, and baggage claim, then probably an hour or so to get to my final destination. If the plane takes zero time, it's probably a four hour trip.

Concorde took about three hours, giving a seven-hour total trip time. Current commercial flights take about eight, giving a twelve-hour flight time. Both are basically 'I'm spending the day travelling'. If you live in London and are getting a helicopter from JFK to your office at the far end then Concorde can let you turn up for a lunch meeting and be home at the end. But for most people, this wasn't the case, Concorde was just a much more expensive way of spending the day travelling.

Contrast that with flights vs ships. It takes a week for a (powered) ship to cross the Atlantic. The difference between spending a day each way travelling and spending a week each way is huge. A one-week holiday or a short business trip is feasible with a one-day travelling time. If it takes two weeks for the round-trip travel then most people can't take a holiday.

This also applies to computing things. The laptop I started doing LLVM development on took an hour and a half for a clean build. It took five minutes to link a debug build. With my last laptop, I'd often build on a big server that could do a clean build in 10 minutes and link a debug build in under ten seconds (a lot of that change came from moving to a better linker). My new (well, three years old now) laptop is able to do a clean build in a bit over ten minutes and normally turn around incremental builds in a few seconds. Going faster won't improve my productivity.

And building LLVM is an outlier. Most people don't do batch processing things that are so computationally expensive. 20 years ago, I was doing a bunch of video editing things and the editing tools had a preview and final-version render mode. Preview rendering typically took a minute or two for large segments or tens of seconds for shorter snippets. That really slowed down the working. A few years later, most of these things moved to pixel shaders and later versions of the tools I was using were able to do the preview instantly. Faster than that is not useful.

That's great for consumers, it's less good for an industry that's focused on growth.

@KarlPettersson @thomasfuchs

Hmm, I don't remember any 486s with 4 MiB of RAM. My 386 came with 4 MiB and 8 MiB was pretty much the standard baseline for 486s (my father's 486 laptop had 8 MiB).

Windows 95's official minimum requirement was 8 MiB, but it was quite painful with under 16 MiB. NT 4 needed 32 to be usable and was amazing with 64 MiB.

@david_chisnall @thomasfuchs Edited the post a bit. At least in Sweden, many 486 computers with 4 MiB RAM were sold around 1994–95: that probably ended with the release of Windows 95.

@david_chisnall @thomasfuchs One may look at IA computer magazines to get an overview of the market at a given time. 4 MiB was perhaps most common in 486 laptops in 1994, but also entry-level desktops could have that configuration. https://archive.org/details/ByteV19N9

@KarlPettersson @thomasfuchs

Hmm, I was skimming, so the first laptop I saw had 8 MiB of RAM but was also advertising enormous battery life. The next system I saw was page 133, which offered the same bundles with 4, 8, 16, and 24 MiB. Typically, when people make listings like that, the cheapest one is there to catch people's attention and the second-cheapest is the one that they expect to sell the most. The next one I found was on page 269 and also had 8 MiB.

My 386 laptop was apparently released in 1991 and had 1 MiB soldered to the board and four SIMM slots that took matched quads, so supported up to 5 MiB total. That was slightly high, most 386s I encountered had 4 MiB. My school computer room was all 386s with 4 MiB (all booting from a floppy and running everything from a file server over Token Ring, which was really slow when 20 machines all tried to launch Windows 3.11 at the same time).

@mmokhi @thomasfuchs

Not quite. Software is very close to a zero-marginal-cost good, but a spoon producer also has a declining marginal cost. Producing 20 spoons is expensive, your cost is totally dominated by the capital costs. Producing a thousand spoons is cheaper. Once you're selling tens of millions of spoon, you've got a factory and paid off the capital costs, so now it's just the operational expenses that dominate.

@thomasfuchs

If I remember correctly, 95 had grammar checking as you write (but the grammar checker was, and remains, terrible). I can't remember new features after that which were particularly useful.

2000 introduced the ability to click anywhere and fill in whitespace, which made people use styles less and generally made output worse.

EDIT: I had a look at Wikipedia's list of new features in Office 2000 and, of those, the only thing that looks like I'd care about it is the new vector and raster graphics program. Which seems to no longer be part of Office.

@david_chisnall @thomasfuchs The amount of new [useful] features in MS Word over time, graphed against the memory size/system demand of the program, would be illuminating. I can't think easily of anything new in word processing that has excited me since we moved from markup based systems like Wordstar to WYSIWYG editors like Word.

@bencourtice @thomasfuchs

Spell checking as you type was definitely limited by CPU and RAM. I think Word 2 kept the dictionary on disk and read individual parts on demand.

The main RAM thing was being able to include images. I remember doing a school report with two scanned photos in ClarisWorks (for Windows) 1 on a 386 with 5 MiB of RAM, and I had to write the text first and then do the layout because it started swapping with both images in the page and would freeze a bit.

There were other things in ClarisWorks 1 that I still miss. A text box in their drawing program had full access to all of the word processor's functionality. A table in either the drawing program or word processor was a spreadsheet, so you could do things like having a total at the bottom row of a table without having to do the calculation externally.