Tech

Software is over: what ends is the business of selling pure features

October 12, 2026

For the last two weeks, I keep hearing strong claims like “Software is over”, “Software is done”, and “Software is dead”. I looked into the background and decided to organize my own thinking.

My answer so far: what ends is the business of selling pure software features. Software itself will keep growing.

Recent developments

These were separate people doing separate things, all at the same time (numbers are from searches as of Oct 12, and details are unverified):

Sep 13: IW4L, a Rust runtime for Modern Warfare 2, goes public.
Around Sep 30: a decompilation of the original Xbox Halo starts running in the browser (TechSpot, Sep 30).
Sep 30 on: open-source versions of Adobe-style apps, called “ArtCraft”, start shipping one after another (e.g. PhotoCraft).
Oct 3 to 8: a reverse engineering tool called “REA” takes off.
Oct 6: a site that runs GTA V in the browser appears, and is taken down within hours (Startup Fortune).

I could find some details on two of them.

ArtCraft: The most popular one is PhotoCraft, a Photoshop clone with about 55k stars (as of Oct 13). Its README says it is “implemented from public specs and observed behaviour only. No proprietary code, shaders or assets”, so the binaries are never read. There are also Illustrator, Premiere, Acrobat, and Office-like apps. The author, Brandon Thomas (CEO of Storyteller.ai, the company behind FakeYou, formerly at Square, 15 years in the industry according to PetaPixel), said “We’re going to reach 100% parity within a month” (TechSpot). On Hacker News he softened it to 99% in months rather than years. He is also the one who said Software is over. His motivation was being “bit by the dark pattern cancellation fee one too many times” (TweakTown), meaning Adobe’s cancellation fee.
REA: An MCP tool that lets an AI agent investigate apps that have no source code. People used to do binary analysis with disassemblers and debuggers. REA hands that job to an AI agent. Under the hood it uses tools like Hopper, Ghidra, and IDA, and it also handles JS/Electron and APKs. It has about 180k stars (as of Oct 13). Its npm package (rea-agents) was downloaded 10 to 50 times a day until Oct 2, then about 35k times on Oct 10. The pitch seems to have moved from security engineers toward app developers. The two famous lines came from people promoting it: Software is done (Matthew Berman, founder of the AI media company Forward Future; the post), and Software is dead, long live software! (International Cyber Digest, a cybersecurity news account on X, operator unknown).

What ends is the business of selling features

I don’t have a problem with “Software is over” when it is about software that sells features. But huge software that has already become infrastructure is a different story. Operating systems, social networks, video services, and search are not going away soon. Their codebases are huge and hard to clone (that may be only a matter of time), they are expensive to maintain, and they hold more than features: data and community.

The ones that get hit are, I think, software that is small today, easy to copy, and has no real strength besides its features. The test I use is: if someone copied the code exactly, what would be left?

When the cost of writing code drops sharply (I feel it every day), a feature stops being a scarce thing that takes months to build. It makes sense to me that the ArtCraft author started from anger about a cancellation fee. It was a revolt against paying forever for features. So what ends is not software in general, but software that sells purely on its features.

For people who build, this is a tailwind

When features get cheaper, the problems software can solve don’t shrink. We find a lot more of them. People building small tools for their own problems is only part of it. I work on the enterprise side now, and companies still have a lot to gain from software. But most of that work goes well beyond writing features. It includes understanding customers’ needs and gathering requirements, including the complicated processes built up over many years, and delivering real value. Areas where software used to be considered too expensive to build or introduce will get a second look.

My guess is that fewer people will work on each product, and there will be many more products. Development moves from a division-of-labor job to a creator’s job.

If you have a software background and want to build products, I think this is a bonus stage. Your knowledge of tech stacks, design, and development methods lets you give the AI better instructions (that may stop mattering someday, but it does for now). I also think abstract thinking and putting ideas into words are very important for instructing AI well. Most engineers I know are good at this. A few are terrible at it, and I’d guess those people will have a hard time.

The cost: the “it runs, so it’s fine” mindset

When it is this easy to get something running, we end up moving forward with the inside being a black box. I see two separate concerns here.

The first is non-functional requirements: security, data retention, permissions, audit logs, and scalability. The screen can show every feature working, while the data is stored in an unexpected place, or the app falls over the moment users grow 10x. “It runs, so it’s fine” can’t see these. But I think this one is relatively solvable. The items to check are mostly known, so checklists for AI to run will probably appear soon. And frontier models, even without any special prompting, now write better code than people in most cases. That is a big change from a year ago.

The second is how much to test. There is no single answer here. Even before AI, it varied by product and company culture. Once a product gets complex, with many states and many UI patterns, exhaustive testing is unrealistic. I expect we will have AI write as many automated tests as possible and reduce how much people check by hand.

For both, the process and rules for how people understand what AI produces will vary by product and company. I take the adventurous side for my personal projects and the safe side at work. In the end, I think we each have to decide how much risk to take, based on the workplace, the product, and the company culture.

That said, I expect a loop of ship, problems everywhere, and better techniques and rules growing from there. It will probably take two to three years before industry standards become visible.

What this made me think about

When I was a student, I got a little interested in security contests and touched binary analysis. Knowing how the underlying technology works makes it easy to catch up on new things like this (maybe I only understand it by feel). It reminded me how much the basics matter.

Looking back at my past projects, I learned a lot from the hard parts. The more AI does for us, the fewer of those experiences we get. Honestly, losing the joy of solving a puzzle and the feeling of building your own little world makes me a bit sad.

Still, I’m lucky to be the kind of software engineer who loves building products, and I’m building all sorts of things with AI these days. Before AI, many of my projects never shipped, or I shipped them and deleted them because maintenance was a pain. Now I can keep improving things after release, and it’s fun. One example is a “one technical explainer a day” app. At first I reviewed everything the AI wrote. Lately it got annoying, and I’ve been merging without reading… Well, it runs, so it’s fine.

Comments (0)

No comments yet. Be the first!