Evolution has shaped intelligence over billions of years. Modern artificial intelligence (AI) now raises a striking possibility: intelligence might begin to direct its own evolution. When AI agents can modify and evaluate their own behavior, two questions arise—how fast can they improve, and how can that improvement remain safe? We propose a theoretical framework that compares Darwinian, Lamarckian, and agent‑directed evolution after an environmental change. Darwinian evolution relies on natural selection, Lamarckian adds inheritable learning, and agent‑directed evolution lets agents generate behavioral variants, evaluate them, and decide how extensively to search. If learning is sufficiently effective, these mechanisms exhibit progressively faster adaptation. Producing more variants accelerates adaptation, yet the optimal search effort depends on learning ability, problem difficulty, and experimentation cost. Crucially, an agent’s ability to steer its own improvement does not guarantee that improvements are judged against the intended objective. When the evaluation target is misspecified, additional search speeds convergence toward the wrong behavior, even as measured performance appears to improve. In recursive self‑improvement such errors can persist across generations, posing serious AI‑safety challenges. Independent evaluation and oversight must keep pace with autonomous AI development. The central challenge is not merely to accelerate the evolution of intelligence, but to preserve humanity’s capacity to guide its direction, limit its speed, and intervene when necessary.
Review: The paper offers a concise hierarchical comparison of three evolutionary paradigms, highlights the trade‑off between search breadth and learning efficacy, and underscores the need for robust external evaluation to prevent objective misalignment during rapid self‑improvement.