[timeattack]

Cool idea, but isn't this kind of old school? I mean, with all the new AI stuff that can learn and guess what we need to do, isn't the whole record and replay thing a bit behind the times? Wouldn't generative AI solve that?

[mgummelt]

We've run some experiments, and the answer is "sort of...sometimes...for some things".

One of our customers, for instance, records macros that upload new users from their Plato tables into their Hubspot CRM, and doing that field mapping is a lot easier in a formal system like a macro than in English.

It seems use cases that can tolerate more ambiguity are more amenable to LLMs, since that's what English is best at.

[timeattack]

I see your point, but consider this - generative AI has shown excellent performance in mapping tasks. I've seen examples where it flawlessly extracts data from markdown format to JSON, so mapping fields should be as nearly as the same? I bet something like GPT-4 can even figure out mapping by itself.

[mgummelt]

Definitely, though the semantics of field mappings often only exist in a user's head, so will need to be explicitly defined.

English prompts will no doubt be a big part of Plato, though. We're still figuring out exactly where the line is between English and traditional UIs. Have you read any good analysis on the subject? Here's one that explores the limits of LLMs: https://magrawala.substack.com/p/unpredictable-black-boxes-a...