It isn’t that simple. There’s a part of it that generates text but it does some things that don’t match the description. It works with embeddings (it can translate very well) and it can be ‘programmed’ (ie prompted) to generate text following rules (eg. concise or verbose, table or JSON) but the text generated contains same information regardless of representation. What really happens within those billions of parameters? Did it learn to model certain tasks? How many parameters are needed to encode a NAND gate using an LLM? Etc.
I’m afraid once you hook up a logic tool like Z3 and teach the llm to use it properly (kind of like bing tries to search) you’ll get something like an idiot savant. Not good. Especially bad once you give it access to the internet and a malicious human.
I’m afraid once you hook up a logic tool like Z3 and teach the llm to use it properly (kind of like bing tries to search) you’ll get something like an idiot savant. Not good. Especially bad once you give it access to the internet and a malicious human.