[avereveard]

Code requires too much precision and is entangled with legal hurdles

The value here is that the llm can act as a knowledge graph were common sense is preloaded on almost every topic, so that the user can add node and edges on the graph in natural language and perform extraction in natural language

And you don't need fine tuning as long as you can fit the topic in their token space, and with gpt4 reaching 32k tokens you can load a huge amount of text and perform queries on it.

That's what makes the tax return example so interesting. The model has already learned a lot of common and uncommon sense so it will not need the instruction on how to process the text or parse the query.

Forget coding, but everything else is great for.