Smaller side model that just prints the "umm sorry I can't do that" message. And they did retrain it at least once - that's why it got so much faster and why th

Smaller side model that just prints the "umm sorry I can't do that" message. And they did retrain it at least once - that's why it got so much faster and why the model chooser has "default" and "legacy" GPT-3.5. Notice that if you hit regenerate it often actually answers the question though.