They keep adding new models but it's a bolt-on on an underlying architecture based on old assumptions that no longer hold for LLMs with emergent abilities like

They keep adding new models but it's a bolt-on on an underlying architecture based on old assumptions that no longer hold for LLMs with emergent abilities like GPT-3.5/4.