If AI isn’t writing 50%+ of your code by the end of the year, you’re moving too slow and you’ll get lapped by the next competitor who does. The bar will be 90%+ soon after (for me it’s over 99%).
Sonnet 4.5 changed the game
There may have been an excuse before, but not since Sonnet 4.5 released. If you try it and think it’s still not good enough, you’re probably using it wrong. And it’ll only get harder to adapt as it evolves, so start now.
There’s been a paradigm shift in leverage. Your codebase isn’t an artisanal snowflake that’s too pure for “AI slop”. Code is no longer the IP since it’s so cheap to generate.
This is more important than it seems because 1) it compounds as AI improves and 2) it takes time for people and projects to adapt. The longer you wait, the harder it is and the further behind you get.
Changing moats
The meaningful moats now are domain expertise (what), taste (how), speed and brand/distribution (e.g network effects, etc.) along with the feedback/iteration loops.
Over the last couple years I’ve heard folks mention the possibility of a $1B solo founder. I now see how it’s possible. It’ll still take a lot of luck (right person, right problem, right time), but I do see how it’s possible.
Not just me
This clicked for me over the last couple weeks as I experimented with a project where AI wrote 99% of the code. I initially used Opus to help with planning/design, which was incredibly expensive and untenable. Then Sonnet 4.5 dropped in the middle of it, which changed the calculus.
Then I saw this, which made it clear it’s not just me AND not just Sonnet 4.5:
Easier to start from scratch
If your company/team isn’t AI native, it’s often easier to just start from scratch than try to adapt an existing project.
I learned that after building or being involved in half a dozen “AI” projects since AI went mainstream in late 2022/early 2023.
Over the last two weeks is the first time I decided to just start completely AI native (beginning with planning). It’s much easier and moves much faster.
Update: Opus 4.5 and Gemini 3 dropped since I originally wrote this and wow, it’s even more true now. If you’re using something like Cursor my go to is to plan with Opus 4.5 and execute with a cheaper model like Gemini Flash 3 or GPT Codex 5.1 Max.
New models will continue to emerge, but the patter is the same: Analyze and plan with the smart/expensive model, execute with a good but cheaper one.
Welcome to the future.
