We're at the inflection point

For programming AI models

This is my code contribution overview of the last year, specifically contributions to Remote’s codebases:

Until ~last year contributing meaningfully to large codebases by people with zero knowledge of them was still very difficult as AI tools wouldn’t take into account the greater architecture and either under-or-overcomplicate things.

Since the release of gpt-5.4 and opus-4.6, and the fantastic work of the people at Cursor and a bunch of other AI labs, this is now changing. I’ve been able to build a completely new complex feature from scratch in Remote.

After creating a first version that worked, one of our amazing engineers Andrea reviewed my work and left a number of comments. One required a rather large refactor (“this should make use of another part of the app”), which normally would be a massive pain and work.
In my case, I just told Claude to “address the comments by Andrea”, which it did.

A few iterations later in which Claude and I fixed bugs and made improvements, Andrea reviewed the work again ..and merged it!

Almost there

I’m very comfortable with developer tools and version management. That still seems to be a requirement for contributions in larger code bases, but it really is the single remaining barrier.

Apps from scratch anyone can now make with Lovable and others. It’s a matter of time before truly everyone can contribute (which was our vision back in 2013 when we started work on GitLab).

The progress over the past months has been so incredibly sharp that my window for confident predictions about a change in AI model capabilities is at best measured in months: I don’t know how much better AI models will be in one year from now, other than that they will be much better.

We’re not at (or near?) AGI just yet. But programming is ~solved.