Thanks, very thought provoking and seems like a pleasant framework .
To help myself digest the implications let me try to distill the statements. I will intentionally make them slightly “stronger” than author’s. I then see if I agree with those.
1. Generalizability is a smooth monotonic function of training data. 2.Current models are not being optimized for generalizability. 3. Training data are bandwidth limited by deployment.
The statement #1 is challenged by grokking. The step-like improvement in generalizability is expected. The question is where is the threshold. The statement #2 seems to have been true up to date. I am aware of at least one startup case https://www.xentlabs.ai/ that recognized this insight and now explicitly optimizes for generalizability. The statement #3 will hold if a) synthetic data are not sufficient for grokking generalizability b) there is no easy-for-deployment industries, that will provide sufficient training data at a low cost. Is SaaS such an industry?
After all this I tend to believe that the road-block is not particularly real. The deployment, partial and targeted on easy industries, will continue to happen and generalizability will grow. In the business context we will have something akin to RLHF but RLManagerF. The “grokking threshold,” is naturally unknown.
Thanks, very thought provoking and seems like a pleasant framework .
To help myself digest the implications let me try to distill the statements. I will intentionally make them slightly “stronger” than author’s. I then see if I agree with those.
1. Generalizability is a smooth monotonic function of training data.
2.Current models are not being optimized for generalizability.
3. Training data are bandwidth limited by deployment.
The statement #1 is challenged by grokking. The step-like improvement in generalizability is expected. The question is where is the threshold.
The statement #2 seems to have been true up to date. I am aware of at least one startup case https://www.xentlabs.ai/ that recognized this insight and now explicitly optimizes for generalizability.
The statement #3 will hold if
a) synthetic data are not sufficient for grokking generalizability
b) there is no easy-for-deployment industries, that will provide sufficient training data at a low cost. Is SaaS such an industry?
After all this I tend to believe that the road-block is not particularly real. The deployment, partial and targeted on easy industries, will continue to happen and generalizability will grow. In the business context we will have something akin to RLHF but RLManagerF.
The “grokking threshold,” is naturally unknown.