What’s the larger issue? It seems like if large ML models are trained in particular data and work a particular way (usually without text prompts), there isn’t any “speaking to machines” translation issue.