First end-to-end trained swarm (twitter.com)

Researchers have unveiled Kardashev-0.7, the world's first trained swarm consisting of 32 distinct AI models. This innovative approach utilizes Reinforcement Learning for Population Scaling (RLPS), enabling the models to develop specializations and complementary capabilities organically. As a result, the swarm achieves remarkable performance while significantly reducing resource usage—operating at only 0.007x to 0.02x of the traditional inference cost and requiring just 0.03x of the memory typically needed for similar tasks. This development is significant for the AI/ML community as it mirrors a fundamental principle of human civilization: the power of specialized collaboration. By showcasing how multiple models can learn to work together and enhance each other's capabilities, the Kardashev-0.7 project presents a new paradigm for scaling intelligence. This shift towards leveraging model count rather than solely enhancing individual model performance could lead to breakthroughs in efficiency and effectiveness in various AI applications, pushing the boundaries of what is achievable in machine learning.