Recursive self-improvement (RSI) — the capacity of an AI system to improve its own performance across successive generations without external retraining or manual intervention — has been widely discussed as a theoretical milestone yet rarely demonstrated empirically. We present FELIX, a population-based artificial organism that achieves reproducible RSI across three successive generations under fixed rules and a staged environmental selection protocol. In five independent runs across five distinct random seeds, every run produced a full positive improvement chain: Gen 0 → Gen 1 → Gen 2, with each generation sustaining measurably higher fitness than the last. The mean improvement from Gen 0 to Gen 2 across all five runs was +129.7%, with a range of +86.3% to +188.0%. No external gradient, hyperparameter change, or code modification occurs between generations; all improvement arises from FELIX's internal dynamics. We present the experimental results, measurement protocol, and limitations of this demonstration. Implementation details are not disclosed at this time.
Junai P. Felix (Wed,) studied this question.