At least Hassabis has an understanding that AGI will require more than an LLM, and understands some of what is missing. He has never spoken publicly about any vision of what an AGI architecture would look like, other than requiring some more "Transformer-level" breakthroughs, so it seems hard to say that he would have failed, other than his 2030 projection seeming unrealistic.
My only criticism of his AGI direction was that he has talked about retaining an LLM as a component of that, but it's hard to tell if that is/was just short-term pragmatism, and a product-based path, or if he really believed this was the best direction. On the face of it having a pre-trained LLM at the heart of an attempt to build a human brain (build true AGI) is an admission you have failed, since if you build a powerful enough (human level) learning architecture it would be able to learn language for itself, not need to have it baked-in. If your version of AGI is not capable of learning language, then what else is it incapable of learning? It would certainly reflect sub-human rather than super-human capability.