4 ms·
I was kinda hoping the article was about how terrible a user interface chat is. All the implementations I've seen so far fall into one of two categories: deter
by specproc 3y ago
I was kinda hoping the article was about how terrible a user interface chat is.
All the implementations I've seen so far fall into one of two categories: deterministic or non-deterministic.
The deterministic ones (e.g. that comically bad Mcdonald's job application bot) operate like those old telephone interfaces, where you need to work your way through a conversation tree to input data. This should be a simple form.
The non-deterministic ones, which are actually powered by LLMs are worse. Here there's no structure, no guarantees, and a fuzzy, stochastic approach to everything. Fine for some things (e.g. coding support or whatever normies do with ChatGPT) but terrible for controlling an application.
- bick_nyers 3y agoYeah for a chatbot to work you really need to force the non-deterministic approach to converge to the actual actions that can be taken. For the experience to be good I think you need to provide training data, such as a user story/feature request (when I ask for X, provide X, fuzz over 1000 different ways of asking for X). Or you can just not do a chatbot.