TODO: Applications enabled by LLM
What if we use LLM... to engineer LLM based applications?
More recent LLM such as llama2 can roleplay a pretty reasonable prompt engineer, and some low grade LLM enabled applications are fairly feasible to be auto-engineered by them.
By LLM enabled applications I mean things that are possible when backed by a general purpose LLM as an AI core, with some helper codes surrounding it allowed. (TODO: link to that training slide)
By low grade I mean those use case that rely on relatively easy (not to look down on them) domain and didn't hit one of the LLM's weak spots. So marketing copy editor yes, full fledged programmer no.
In these case automation is actually possible too, as well as a more interactive, chat based approach. For automation we can use simple prompt-chaining.
- User specify topic plus instruction/requirements. If no idea, can also choose a broad domain and let AI generate app idea.
- AI design prompts and examples, and then auto-extarct/format it into a machine readable format, something like an Application definition. Here we use JSON.
- User can test drive the app by "Deploying" it live, and then do a sample run.
Currently it includes the following:
meta: General info like app name, description, and the main text prompt.ui: Programmatic UI layout for a simple form with (dynamic) fields so that user can input info more flexibly.eg: A set of AI generated example inputs data. This allow end user to get started quickly.chain: Simple prompt chains with a set of followup questions that AI can automatically generate answers to in batch mode.