Essential AI's Aleksa Gordić reveals MAI-Thinking-1 training details, including tied embedding weights and disabled NVLink SHARP

few surprising details about MAI-Thinking-1: 1. AdamW 2. input and output embedding weights are tied - from ablations i've seen so far this only made sense for ~4B and below 3. NVLink SHARP disabled to ensure determinism but at the cost of reduced performance