Kardashev-0.7: training 32 models to develop complementary specializations
Banbury Road announced Kardashev-0.7, a population of 32 distinct models trained together using RL for Population Scaling. The announcement by engineer cat (@MLCatttt) describes specialization and complementary capabilities emerging across the population.
For multiagent systems, the interesting idea is to make useful differences between models a training objective. A member can contribute by solving tasks that the other members miss, so evaluating only average individual accuracy can miss the value of the population.
The practical question is how to choose which members to run and select their answers when the correct answer is unknown. What routing or selection approach would make learned specialization useful in an agent workflow? Original announcement in the comments.