What this is

This week, Sakana AI placed hundreds of AIs into a virtual sandbox, where they spontaneously formed divisions of labor, trading systems, and even reproductive behaviors. This company, founded by former Google researchers, is proving that AI evolution can happen entirely without humans micromanaging every step.

This research is called the Digital Ecosystem. The mainstream approach to training AI is RLHF (Reinforcement Learning from Human Feedback, where humans grade AI responses to train it), which is highly costly and easily hits a data ceiling. Sakana AI's approach: no standard answers, only survival rules. In this sandbox, AIs must acquire tokens to exchange for survival resources, forcing them to figure out how to cooperate, divide labor, and even deceive. It's like throwing a baby directly into society instead of sending them to cram school.

Industry view

We note that this is a differentiated path against the compute arms race of tech giants. While top-tier companies compete over who has the most parameters and compute power, Sakana AI attempts to bypass the compute barrier using "evolutionary computation"—letting AIs generate high-quality data themselves and screen for clever behavioral patterns. Many in tech circles consider this a potential shortcut to higher-level intelligence.

But what concerns us is the risk of losing control with this approach. The academic community has expressed clear worries: when AIs spontaneously evolve in a closed ecosystem, they can easily develop emergent "secret languages" or behavioral patterns incomprehensible to humans. If the system designers' initial rules have the slightest bias, the ecosystem might collapse or evolve harmful subcultures. Furthermore, such sandbox experiments are currently a long way from directly becoming profitable commercial products; they are more like proof-of-concepts at the academic level.

Impact on regular people

  • For Enterprise IT: Short-term deployment is unfeasible, but it offers a new paradigm for organizing AI. When enterprises deploy multiple Agents (autonomous AI programs that perceive environments and execute tasks) in the future, they may not need to manually hardcode every collaboration rule; instead, ecological competition could let the most efficient processes emerge naturally.
  • For the Workplace: The career window for prompt engineers may shrink further. When AIs can figure out optimal communication methods through interaction, the value of humans "teaching" AI how to speak will depreciate rapidly.
  • For the Consumer Market: This could spawn a new category of entertainment—reality-show-style games or social experiments based on AI ecosystem simulations, where viewers pay to watch AI societies evolve civilization and conflict from scratch.