Researchers have developed an AI model called EMO that achieves nearly full performance using only 12.5% of its experts, focusing on content domains rather than word types. This advancement could make mixture-of-experts models more viable for use in memory-constrained environments.