Banbury Road trains 32 models together, with the models independently developing distinct areas of expertise.

Beating AI Express News: AI lab Banbury Road has released Kardashev-0.7, a system composed of 32 distinct models. The team trained these models together via reinforcement learning, without predefining each model’s specific tasks, enabling them to develop unique specialties and complementary capabilities during training. This methodology is called RL for Population Scaling (RLPS). Prior RLPS experiments only reached up to 16 models. In the 8-model trial, the jointly trained model population scored 81.70%, outperforming 71.04% for 8 independently trained models and 72.65% for the same model sampled 8 times. In the 16-model experiment, 22 questions were answered correctly by only one model each, verifying that different models did learn distinct capabilities. Kardashev-0.7 expands this to 32 models. Banbury Road claims it delivers state-of-the-art performance at 0.007 to 0.02x inference cost and 0.03x memory usage. Its API beta is currently on a waitlist. However, the population scores from the earlier 8 and 16-model trials were derived post-hoc by picking the best result from multiple model outputs using standard answers. Since real-world applications lack standard answers, Banbury Road has not yet disclosed a full solution for the system to automatically select correct responses.