The context window is comparable to human short-term memory. LLMs are missing episodic memory and means to migrate knowledge between the different layers and into its weights.
Math is mostly impeded by the tokenization, but it would still make more sense to adapt them to use RAG to process questions that are clearly calculations or chains of logical inference. With proper prompt engineering, they can process the latter though, and deviating from strictly logical reasoning is sometimes exactly what we want.
The ability to reset the text and to change that history is a powerful tool! It can make the model roleplay and even help circumvent alignment.
I think that LLMs could one day serve as the language center of an AGI.
Math is mostly impeded by the tokenization, but it would still make more sense to adapt them to use RAG to process questions that are clearly calculations or chains of logical inference. With proper prompt engineering, they can process the latter though, and deviating from strictly logical reasoning is sometimes exactly what we want.
The ability to reset the text and to change that history is a powerful tool! It can make the model roleplay and even help circumvent alignment.
I think that LLMs could one day serve as the language center of an AGI.