The end of software

Tue Sep 13 2022

It seems reasonable that, in under 5 years (85% confidence), one will be able to write a reasonable description in english and generate the entire codebase from AI. Give a PR title & description, and out comes the PR. Maybe the PR itself is reviewed by AI or maybe it just runs because it's so often correct.

This might mean the end of software in the way we think of it today.

The "iPhone" of the future will not have an App Store.1 If there are "apps", they'll be generated on the fly, change based on your behavior, and be beautifully designed to match your exact tastes. iPhone is a cash cow so it seems hard to imagine Apple actually releasing this. But whoever2 does has a remarkable advantage over Apple: every single app any consumer could ever want exists on AIphone. No need to get Meta to create a WhatsApp for AIPhone; it makes one on the fly. Every app is free or of enormously low cost. There's no ads and every app works perfectly on your model. Every app works great with every other app, every app is perfectly accessible and can redesign itself if you need a bigger text size.

The future of the computer is the generative interface, not apps.


What happens in the meantime?

Software likely becomes easier to build. Teams can become much smaller and there will be far more of them. Players with a lot of existing distribution might win, but probably not in all cases. Some developer tooling companies might struggle. New tooling products might disrupt them by allowing folks who previously did not "code" to "code". Models and runtimes that make it safe to run the code generated by an AI and understand what it's doing may turn into interesting products, or they may be incorporated as features of other products.

Venture may have a difficult time investing in the companies they're used to investing in, but probably have several years more of returns on the traditional strategy. Youtubers have successfully "beat" TV to some extent, but it's rare that a Youtuber raises funding. It's feasible the same thing happens to software.

  1. Maybe 10 years with 30% confidence, 15 years with 65% confidence. This depends more on powerful on-device, low-power GPUs being usable; hardware seems to move slower than scaling training.
  2. My bet is OpenAI makes AIPhone. They're far ahead in the AI, have captured a reasonable pool of talent, and have a strong partnership with Microsoft who does not have a mobile app store that would be threatened, and could feasibly pull off the hardware side.