So where are all the AI apps?

What's sauce for the goose is sauce for the gander. If you make that argument that 'I don't believe in kinks or discontinuities in code release due to AI, because so many AI coding systems have come out incrementally since 2020', then OP does provide strong evidence for an AI acceleration - the smooth exponential!

I think this represents a fundamental misunderstanding of how these AI tools are used most effectively: not to write software but to directly solve the problem you were going to solve with software.

I used to not understand this and agreed with the "where is all the shovelware" comments, but now I've realized the real shift is not from automating software creation, but replacing the need for it in the first place.

It's clear that we're still awhile away from this being really understood and exploited. People are still confusingly building webapps that aren't necessary. Here's two, somewhat related, examples I've come across (I spend a lot of time on image/video generation in my free time): A web service that automatically creates "headshots" for you, and another that will automatically create TikTok videos for you.

I have bespoke AI versions of both of these I built myself in an afternoon, running locally, creating content for prices that simply can't be matched by anyone trying to build a SaaS company out of these ideas.

What people are thinking: "I know, I can use AI to build a SaaS startup the sells content!" But building a SaaS company still requires real software since it has to scale to multiple users and use cases. What more and more people are realizing is "I can created the content for basically free on my desktop, now I need to figure out how to leverage that content". I still haven't cracked the code for creating a rockstar TikTok channel, but it's not because I'm blocked on the content end.

Similarly I'm starting to see that we're still not thinking about how to reorganize software teams to maximally exploit AI. Right now I see lots of engineers doing software the old way with an AI powered exo-skeleton. We know what this results in: massive PRs that clog up the whole process, and small bugs that creep up later. So long as we divide labor into hyper focused roles, this will persist. What I'm increasingly seeing is that to leverage AI properly we need to re-think how these roles actually work, since now one person can be much responsible for a much larger surface area rather than just doing one thing (arguably) faster.

"The first 90% of the code accounts for the first 9% of the development time. The remaining 10% of the code accounts for the other 9000% of the development time"

A lot of the real value shows up as workflow compression instead. Internal tools, one-off automations, bespoke research flows, coding helpers, things that would never have justified becoming a product in the first place.