I think maybe, but I wouldn't go about it like this.
I'd have it build an internal DSL on top of a preexisting language that millions of dollars of data science and compute have already considered.
The "builder pattern" is probably the most capable path I've seen out of LLMs. These models are really good at building type hierarchies that support this. They're also really good at using these types given some examples.
Other DSL paths like SQL are more marginal in my experience. Terse, declarative languages seem to work poorly unless we are using something verbatim from the training data.
Having the LLM propose a whole language feels like starting over from zero unless having it create arbitrary toy languages is the end goal.
I'd have it build an internal DSL on top of a preexisting language that millions of dollars of data science and compute have already considered.
The "builder pattern" is probably the most capable path I've seen out of LLMs. These models are really good at building type hierarchies that support this. They're also really good at using these types given some examples.
Other DSL paths like SQL are more marginal in my experience. Terse, declarative languages seem to work poorly unless we are using something verbatim from the training data.
Having the LLM propose a whole language feels like starting over from zero unless having it create arbitrary toy languages is the end goal.