I'm really excited for where the open weight models go from here. I've had fun with just CPU inference on old servers that only have AVX1; here's hoping for commoditized TPU-like hardware!