Well this preprint’s title is already misleading because there is no evidence of “reasoning” in these models and even small changes in the benchmarks break and semblance of “reasoning.”

→ View original post on X — @timnitgebru