> Accuracy issues will improve. Mostly likely the improvement necessary for LLMs to answer a vast majority of queries as reliably as necessary will happen quickly. Not for all cases. But for a sufficiently large set of cases to make Google alternatives very compelling.
On what basis do you say this? Is it anything more than your gut feeling?
How much additional training data will be needed in order to achieve these improvements and do you have a concrete reason for thinking that amount will be sufficient? Do you have any expectations as to how this training data will be obtained?
It’s only a gut feeling, based largely on the predictions and accounts I’ve seen from those with actual first hand knowledge around the reasonable questions that you pose.
My framework for thinking about this at a technical level is that short term improvements can come from a number of factors, including optimizing the number of parameters, more data, better data preparation and selection, hyper parameter tuning, etc. I understand that many of the emergent capabilities we’ve seen recently from LLMs have come from more data, and I’m honestly not sure how to assess where short term accuracy improvement is most likely to come from.
On what basis do you say this? Is it anything more than your gut feeling?
How much additional training data will be needed in order to achieve these improvements and do you have a concrete reason for thinking that amount will be sufficient? Do you have any expectations as to how this training data will be obtained?