From the other examples, this doesn't seem like it's even an LLM, it's more 2010's level chatbots.
it’s terrible if it’s an LLM, that’s for sure.
my guess: they took an off the shelf open source model, plugged in the loc’d copy for all the websites and told it to build questions against that corpus
it also doesn’t seem to handle things like region specific requests either. embarrassingly bad
More likely it's pre-LLM techno. Citibot has been around since at least 2017, no-one had a LLM chatbot back then, but pattern-matching and algorithms.
May 10, 2025 05:56
