If you're OpenAI you're going to look at all the training data which is available from people having used your LLM. If you ask it a question, it answers and you then ask it 20 more times to refine it then that's a massive amount of quality data for OpenAI. And the whole world is basically doing that, so they're going to have a massive amount of people who are asking the same questions and going through the same refinement processes to qualify their answers. This is going to lead to a good refinement of how those questions need to be answered going forward when it's compared to the datamodeling they already have.

It's going to be hard for competitors of course, but that's a different story.

Either by encoding a morse code into output by occasionally selecting suboptimal candidates, or by using a second AI trained to recognize AI. Neither method is known to work well.

If I were OpenAI I’d look in my logs.

I'm guessing that training an LLM to recognize LLM output should be an easy task.

If it's so easy, you could stand to make a lot of money bringing a quality product to early market.

Easy for those who already know how to train an LLM, I mean.