Dario Amodei likes to give bold predictions in the spring. Last year it was that, within 12 months, AI would be able to write 100% of all code. I didn’t believe him and I was wrong. It only took one step change to get from where we were then to where we ended up in November 2025.

His new prediction is that in 6-12 months it will be able to do the entire end-to-end software development workflow, without any of the guidance and judgement that is currently still required by humans.

Goodbye! 👋🏻

I believe him this time because I do think that capability is one step change away and I have no reason to believe we are at the top of the s-curve.

Fundamentally: the difference between the output I get from Claude Code and the final code I ship is guidance that I supply to Claude Code through text on my computer. Input-output-input-output-input-output. Any consistent input-output data set on a computer is a curve that models can be fit too. I have a finite amount of guidance I can supply.

I have laboriously documented everything I know, care about, and believe to be true when it comes to software development and Apple platform development into agent skills. Claude Code and Codex, with small loops, can use them nearly perfectly and produce perfect-to-me output without any guidance. A child could sit down at my computer and get better output than I could have manually produced 2 years ago. With one step-change improvement to the models there is no way I will be necessary at all.
@kyle Can my son borrow your computer?
@thillsman your son is an existential threat to me
@kyle @thillsman Tyler probably shouldn’t have named him Claude.