Comment by luma

Where RL can play into post training there's something of an anti-moat. Maybe a "tow rope"?

Let's say OAI releases some great new model. The moment it becomes available via API, everyone else can make use of that model to create high-quality RL training data, which can then be used to make their models perform better.

The very act of making an AI model commercially available is the same act which allows your competitors to pull themselves closer to you.