Only the GFG models know anything internal. Regular Gemini isn't trained on any of that. And GFG is a much older base model, so people use the regular one. If the tools seem to handle google3 code ok, it's only because of skills and not the model itself, and then you run into issues with skill bloat. Sometimes the A/B test would give me the bad model of the day that'd try to grep all of piper.

Start in a blank directory and tell it to spin up a boq Scaffolding stubby server that responds with "hello world." Unless something has changed after I quit a few months ago, it won't know how to do that locally, let alone actually deploy it. Try the same outside Google with like a Flask server on AWS or GCP.