DeepMind has developed a new approach called streaming DiLoCo that makes distributed AI model training 100 times more bandwidth-efficient without compromising quality. Chinese scientists have demonstrated that open weight AI models can self-replicate, highlighting global safety concerns. Facebook has introduced MILS, a new zero-training method to improve the performance of large language models using a generative adversarial network-like approach. Researchers are expressing concerns that even incremental AI advancements could undermine human control over critical systems. Additionally, a new Chinese AI reasoning model called Kimu k1.5 shows promising performance but is not open source.
Sort: