Since GPT-1, we have known that large-scale data drives AI progress. What is less clear is how much targeted data is still needed today. If we want a model to become good at task X, will it work out of the box? Can we just apply RL? Do we need SFT traces? Or do we still need humans on platforms like Mercor to create expert data?
Since GPT-1, we have known that large-scale data drives AI progress. What is less clear is how much targeted data is still needed today. If we want a model to become good at task X, will it work out of the box? Can we just apply RL? Do we need SFT traces? Or do we still need humans on platforms like Mercor to create expert data?