A 300-person Chinese startup, Moonshot AI, reportedly shipped a model comparable to Opus 4.8, despite being compute-poor, according to a DeepMind researcher. The researcher argues that training is efficiency-compressible through techniques like MoE routing, INT4-native quantization, and better data curation, enabling smaller labs to create frontier models.
The compute moat is weakening as efficiency gains allow smaller, compute-constrained labs to reach frontier model capabilities, challenging the effectiveness of export controls.