AI 'Babies' Now Inherit Traits From Two Parents, Study Finds
Could machines one day evolve on their own? This small experiment hints at how.
Most AI today learns the same way: humans feed it huge amounts of data and it slowly gets better. This study tried something different. A researcher named Lyes Saad Saoud built a simulated 'newborn' AI — he calls it a Machine Zygote — that never learns at all. Instead, two 'parent' programs have their internal instructions randomly altered and mixed together, like shuffling two decks of cards. That mix becomes the newborn's built-in instincts, which it is then tested on. The paper uses the word 'germline' for the instructions that get passed down, and 'soma' for the body those instructions build.
The results were surprisingly clean. He ran a breeding grid of 640 virtual offspring and found that five of six behaviors — including how fast they moved and how often they changed their walking rhythm — depended on which two parents they came from. Roughly 36 to 53 percent of the differences between offspring traced back to parentage. He then ran a sharper test: swap in a different parent's instructions while keeping everything else identical, and the offspring changed in a predictable way. Some offspring even outperformed both parents.
Now the honest part. This is a computer simulation, not robots and not biology. The paper explicitly says it does not prove physical inheritance, real genetics, or that machines can evolve on their own. It also tested whether the simulated 'growing up' process mattered, and it didn't — the same results showed up without it. And it's an arXiv preprint, meaning other scientists haven't reviewed it yet. There is no product here, nothing to buy, and no money saved today.
So why care? Because it's a careful, reproducible test bench for a big question: could machines someday improve by combining successful designs rather than by humans rewriting code line by line? That idea matters for robots, which need to adapt to messy real-world conditions. This paper doesn't get us there. But it sets up a way to tell real inheritance apart from coincidence — which is the unglamorous work that makes later breakthroughs possible.
- A simulated AI 'baby' inherited traits like speed and walking rhythm from two parent programs — without any training at all.
- In a breeding grid of 640 virtual offspring, roughly 36–53% of behavioral differences traced back to which 'parents' it had.
- The paper admits its limits: it's a simulation, not biology or robots, and it has not yet been reviewed by other scientists.
Why It Matters
Hints at machines that improve by 'breeding' instead of human coding — promising for robots, years from real use.