There is no such thing as the best in software. Software companies converge on a plateau defined by the market’s diminishing marginal utility. It sucks a little bit for everyone, but it’s still worth it for purchasing. All software begins to solve a problem for one, delights a few along the way, grows by consensus, and dies from frustration, or fades into a cultural leap. It’s been pretty rough to keep the party going from the peak. A slow death by consensus led to the eventual demise of many products, such as Skype, Yahoo Answers, and Orkut.
There was such a thing as free software, like free beer of a magnitude never before seen on the planet. These were high-quality things. I grew up in a generation of software that was free by hook or crook. Where I grew up, we didn’t have the money to pay. Most people I knew installed Microsoft Windows on their desktops, and some ran Ubuntu. Practically no one I knew growing up purchased a Microsoft Windows license for their desktop. Microsoft Windows of the 90s is a monumental engineering artifact even by today’s standards. They pretty much had to give it away for free to entire countries and still become the most valuable company in the world.
The internet enabled software access across devices, and the results, in cases such as photo sharing, have changed our culture. The internet was free, too. Growing up, the broadband connection and the computer were the most costly items to connect to other people through wires. The internet remained free throughout the rise in hardware capabilities, from a 1MB floppy disk to a 1TB stick you could accidentally lose in your pocket.
Multi-tasking on computers is now a muscle memory and a hazard that doctors need to warn their patients about the adverse effects. The internet grew from not being able to reliably send a byte across the Atlantic to streaming a 10GB movie on YouTube over the weekend and reading emails on the computer at the same time. It remained (sorta) free throughout. I can’t actually believe Google Maps remained free. It’s become a pretty much essential good to navigate, and there’s been no good enough alternative for a really long time. They had to be free because of the consumer sentiment against purchasing software.
Where I grew up, during school, most people had absolutely no conception of what software was. They’ve never seen one, though they’ve read about it in books. Those who came across it saw it in a dusty back-office room, and they were lucky if it worked; they couldn’t imagine how to get people to pay them over the internet instead of hand-delivering a check. Shops in the neighborhood rented computers. I saw a computer in my secondary school. I was told about it from a distance, and it seemed high-maintenance compared to TV. I really didn’t imagine computers morphing my life in various ways till I joined undergrad, and by then it was pretty clear even to my previous generation. Probably for the first time, the masses paid for WhatsApp before Meta acquired it. I never really paid for software till the 2020s. It’s pretty cool they pulled off this thing.
It certainly was easier to directly ask enterprises for the money they needed for the software. There was always a free and open source software looming over them, about to eat their share of the pie. Enterprises always wanted changes more frequently than FOSS could ever deliver, which meant there was money to be made doing that thing, and it became their moat. Even now, their moat is their Achilles heel. The more complex the software is, the harder it is for people to understand, creating opportunities for competitors to lower the accessibility barriers of existing products. These are usually simplified and narrower versions of an existing, much broader offering. Each fragmentation creates more failure points that hide under the rug. More people needed to work, but significantly fewer than what Henry Ford needed to build cars.
Interestingly, free and open-source software dominated in some areas. Even for people who didn’t know how to compile source code into a binary, there were straightforward entry points. They suffered from a lack of thickness in their labor-market opportunities, but made up for it with strong interest and obsession. After all, a thin but consistent lifeline has some advantage over random sugar highs. With some exceptions, such as MATLAB, open source has delivered better, dazzling software like Python. Matplotlib was such a great open-source graphing project. Their compatibility with MathWorks’ API made it more complicated than necessary, leading to several higher-level abstractions. But all abstractions leak.
The dark side of the software industry is its integration software. The thing that sits between two things built by two different companies. In the case of Windows OS, Microsoft had to partner explicitly with every single hardware device manufacturer to ensure its software actually starts up without throwing up the customer who just paid them (or didn’t). Microsoft has an incredible legacy of its software being backward-compatible for decades, until Satya began to focus more on market cap and social media moments. I didn’t realize that’s what he meant when he published his book Hitting Refresh. I guess he meant it’s not going to work even after a refresh. It’s incredibly hard to consistently write backward-compatible bug-free code that lives in a remote, unobservable environment for the remainder of its life, and they did.
The exclusive contracts between vendors really made Linux adoption harder because people didn’t want to write the integration software for Linux distributions and hardware. That meant Ubuntu would not work on the latest laptop, even after Lenovo gave its support. It gets complicated when multiple vendors are involved in installing new applications. You’ll be living through power/energy regressions, graphics cards that don’t work, printers that don’t work, and slide decks that don’t work on the day you need to deliver a presentation to a potential customer.
Then there’s the mother of all integration gigs, like the Enterprise Resource Planning that SAP software leads at the moment for businesses that absolutely must run correctly to keep the world from chaos. An integration project to connect a large payment point-of-sale system across continents due to a merger is a five- to ten-year project costing millions and potentially failing, like the Apollo mission.
But now, this dynamic seems to be changing. Pretty much every AI-native product I download wants me to pay $20/mo after a week of use. This has implications in the order of leasing a new Toyota Corolla. The sassy old ones slapping copilots on the side are bundling their next best features behind the paid Gemini-powered chat window. There’s so much demand that Google is rate-limiting the API calls on its Gemini flagship models. The company with the best infrastructure on the planet and free cash flow in the checking account of some nations is rate-limited in showing more revenue in its quarterly reports. Maybe the clunky ChatGPT is actually getting enough of the small things done that fell through the cracks of the fragmented, clunkier SaaS market, so that marginal utility is higher.
I use about a dozen software programs every day. Email, notes, spreadsheet, documents, photos, maps, messaging, music, etc, are obvious ones by now. Software is going to cost more than electricity if everyone charges me $30/mo individually. Productivity apps like Google Workspace are bundled and priced based on average usage, and the total might reach $100/mo if that pattern expands more broadly. That’s still as much as electricity, and I think, at least in the US, consumption patterns seem to favor this trend.
I’ll say it again, but differently. Integrations break software, and the costs shape entire industries, and it doesn’t have to be about integrating between two different companies. Two teams at a big corporation show the same behaviors, and probably even worse. Sometimes it’s much easier for Google Cloud to acquire a company than to merge its Developer Console with the Google Cloud Console, even though people are clearly confused about it. They recently made it happen, yay! But looks like their insane page load times, more than twice that of their next competitor, have not yet been registered on their priority list. It’s actually not that simple because Google Cloud has an internally consistent worldview, thinking it’s a good idea to wrap several billion-dollar product suites serving millions of users into a single website. We are talking about systems within Google Cloud that were individually evolved by different people following different patterns, and now have to dance in concert at runtime. Google Cloud’s core value proposition stems from its approach to parameterize accessibility, which differs from Amazon Web Services. Some examples include Google Cloud, which defines IAM by user roles, but AWS defines it by service; The Cloud SQL offering is similarly differentiated around the accessibility of DB administration. On the internet, websites can update their software whenever they choose, unlike Microsoft. So, Google had it easier than Microsoft at least on the backward-compatibility crisis, but it’s still pretty hard.
Integrations are at the core of the problems of human accessibility. Humans have limitations in their ability to perceive complex information within constrained efficiency and form factors. There are platforms with general-purpose APIs. Several others were built on top of them, with higher-level abstractions for a more specific audience, making it easier to access their use cases.
As much as we don’t want to believe, interface design and development remain challenging, even now, limited by the form factors available. The corporation wants you to use Microsoft Teams so that it integrates well with the rest of the infrastructure, although you’d prefer Slack, but you’d actually prefer something else.
Accessibility is fundamentally constrained by the mechanics of the interface. If the list of groceries you want to display does not fit on the checkout screen, there needs to be a scroll and hardware to navigate. You’ll need differently tuned interfaces depending upon user motivation. If each of those items has a lot of information to display, and the primary purpose of consumers is to compare toward a purchase vs. monitor correctness, the first probably doesn’t need a side panel, but the other definitely is better off with a list on a side panel. These complexities are less about a lack of intelligence and more about fundamental constraints of the interface. Sure, there are improvements to be made. These have downstream effects on integrations within the interface development.
The conceptual framework for solving the accessibility problem is available in current software in the form of extensions, such as Chrome Extensions, Firefox Plugins, and VS Code Extensions. The previous generation of software, notably Emacs and Vim, had outperformed in terms of extensibility despite accessibility. The most interesting event in software, driven by large language models, is a shift along the accessibility axis. Users are increasingly paying for it now, unlike before. Interestingly, the total cost of creating a more perfect utility for a narrower task is actually decreasing, along the axis of people assembling their own experiences.