AI Builds Itself: Recursive Self-Improvement in 2026 by H. Peter Alesso
The machines are already improving themselves. You just have not been told how fast.
In the spring of 2026, Claude began writing between seventy and ninety percent of the code used to train its own next version. AlphaEvolve spent over a year optimizing the training process for the very models that powered it, recovering enough wasted compute across Google’s global fleet to power a small country’s worth of servers. OpenAI’s Codex debugged the training pipeline that produced it. And a 630-line Python script showed that anyone with a single GPU could run autonomous AI research experiments overnight.
Buy the book, and follow the author on social media:
Get This Action & Adventure Book.
Learn more about the writer. Visit the Author’s Website.
Author Bio:
I love words, but that wasn’t always true. I grew up with a talent for numbers, leading me to follow a different path. I went to Annapolis and MIT and became a nuclear physicist at Lawrence Livermore National Laboratory.