Nested transformers train and use multiple submodels simultaneously

Original: this work is super cool, with these nested models you can actually do lots of neat tricks e.g. dynamic routing between submodels, self-speculative decoding, and more! please check out avi's poster at

Source: x.com ↗