"We showed that transformers can execute programs efficiently inside their own inference loop, not as an external tool, but as part of the model itself. This opens a path toward AI systems that integrate learned representations with compiled algorithms inside a single

→ View original post on X — @linusekenstam