New RECLAIM framework grows AI through evolution, not optimization

A new paper on arXiv introduces RECLAIM (Recursive, Ecological, Cognitive, Lifelike, Adaptive, Intelligent Machine), a radical departure from today's dominant AI paradigm. Author Ata G. Zare argues that top-down optimization via gradient descent and RLHF creates inherent structural failures—hallucination, sycophancy, reward hacking—that cannot be fixed with better engineering. The OMEGA shift (Optimization to Emergence through Generative Autopoiesis) proposes cultivating intelligence through computational ecology instead.

RECLAIM rests on four pillars: General Darwinism replaces gradients with blind variation and selective retention; non-agentic emergence substitutes reward functions with environmental physics to prevent specification gaming; the Polya-Hebbian bridge uses urn dynamics to drive specialization; and the free energy principle is reinterpreted as environmental thermodynamics rather than an agent objective. The architecture features autopoietic units bounded by Markov blankets competing for finite computational energy within cognitive food chains and Red Queen dynamics. Zare suggests this setup could spontaneously yield dual-process cognition, sensory specialization, analogical reasoning, and intrinsic motivation—without explicit programming.