Many in the AI community argue that "el LLM es incapaz de ideas originales," and they may well be right. A single LLM is unlikely to be the direct source of truly original breakthroughs. However, transformative ideas are far more likely to emerge from massive, high-density social interactions across vast networks of AI agents. This mirrors classic emergent phenomena seen in complex systems: a single ant is limited, yet ant colonies solve sophisticated problems; an individual neuron possesses little intelligence, yet billions working together give rise to consciousness. Within these agent societies, a dynamic "idea marketplace" naturally forms, where strong concepts are amplified, weak ones are discarded, and diversity thrives through agents with varied personalities, specialized knowledge slices, and distinct objective functions. Over sustained interactions, this leads to cultural evolution, accumulating knowledge that extends well beyond any individual training data. When combined with tool use, real-world feedback loops, and continuous self-improvement, the results become exceptionally powerful. This "AI society" style of interaction is vastly more potent than prompting a single model dozens of times. It closely resembles the way human civilization advances—through discussion, debate, and iteration—but at a speed and scale thousands or even tens of thousands of times greater than humanity alone.

C@C

(view original post →)