[dogma1138]

The value right now is learning which prompts work, which strategies work, what is the sweet spot in terms of model size, is it better to run a model that costs 10 times more per token vs smaller inferior models that might need have 3 attempts at it, what are the best RAG and other enrichment strategies etc.

You can only really get that when you have a large market share.

When your IDE makes a call to OpenAI they get to see only half the picture they don’t know how the user reacts to the output.

So there is a lot of value in being in the client loop also and seeing the full picture.

[paxys]

You know who else has that data, that too for free? OpenAI.

[dogma1138]

OpenAI only sees half the picture, that’s why they want to be in the IDE also so they can get all the usage metrics from the user viewpoint.