3 ms·

> interfacing external systems to an LLM isn’t particularly challenging Based on what do you say that?

• dragonwriter 4y ago
  > Based on what do you say that? The research on both training and prompting such actions, and actually messing with implementation of the latter. (And, by “not hard”, I don’t mean “there wasn’t considerable work to figure it out”, but “its a tolerably solved problem with plug-and-go solutions that Google was among the players doing the foundational research for with earlier models”; Bard doesn't do it now by choice, not Google facing any barrier to doing it.)
  • epups 4y ago
    I'm not so sure. If it was that easy, we would also start seeing it in open source models, and Bard itself I can hardly believe is choosing not to do that.
    • dragonwriter 4y ago
      > If it was that easy, we would also start seeing it in open source models. Its mostly being done on the prompt side, either via ad hoc client code [0] or frameworks like langchain, which supports a variety of underlying models. [1] [0] Here’s one that was posted to HN wrapping the ChatGPT API before ChatGPT announced plugins, but there’s nothing special about ChatGPT to it: https://til.simonwillison.net/llms/python-react-pattern https://til.simonwillison.net/llms/python-react-pattern [1] https://langchain.readthedocs.io/en/latest/modules/agents/getting_started.html https://langchain.readthedocs.io/en/latest/modules/agents/ge...
    • ShamelessC 4y ago
      It is trivial- a matter of prompting and reading API documentation. The latter is well understood by existing developers.