One of the questions I've been asking myself recently is how AI training will evolve when simply adding more data provides diminishing returns.
We've made tremendous progress in scaling up generation of synthetic examples, but it doesn't always equal diversity in capabilities learned. It's possible to generate thousands of different examples which train your model in the same manner.
This is why the data for post-training becomes really interesting. The valuable examples might be the ones which reveal the weakness of the model, which are based on realistic tasks and provide some way to check if the model managed to complete the task.
While searching for such examples, I discovered Parsewave. Their area of expertise is post-training data on engineering tasks, evaluations and traces. But what is interesting is their concept itself - deliberately generating the data on the capabilities which remain challenging for the model instead of generating the big datasets.
What do you think about the future direction of AI training?
Will the future of AI be about generating the massive datasets or becoming really good at identifying a small number of truly useful examples?
[link] [comments]