Their (and Muskrat and others’) quest to build one-model-for-everything has resulted in less reliable systems than what we’ve had before, even in the well defined task of ASR. As in, historically “hallucinations” weren’t problems in ASR systems!

→ View original post on X — @timnitgebru