← Back to context

Comment by TeMPOraL

1 day ago

Every runtime executing LLMs with support for tools does it, starting with the first update to ChatGPT app/webapp that made use of the earliest version of "function calling"? Even earlier, there were third-party runtimes/apps (including scripts people made for themselves), that used OpenAI models via API with a prompt teaching LLM a syntax it can use to "shell out", which the runtime would scan for.

If you tell a model it can use some syntax, e.g. `:: foo(arg1, arg2) ::`, to cause the runtime to call an API, and then, based on the context of the conversation, the model outputs `:: get_current_weather("Poland/Warsaw")`, that is "generating code on the fly to all APIs". How `:: get_current_weather("Poland/Warsaw")` gets turned into a bunch of cURL invocations against e.g. OpenWeather API, is an implementation detail of the runtime.