I would expect research along this way to pick up quite a bit if we confirm the pretrain stage is not scaling as previous expected, thus the scale and architecture would be more stable in the near future, especially if the focus shifts to inference time scaling.