![[src + arch inside] Pausing AI Developments Isn't Enough. We Need to Shut it All Down [1075x2057 - 1.05MB]](https://pic8.co/sh/4KRGz3.jpg){kind=link}
I assume
In this case, you assume incorrectly.
pure computational power, measurable access to amount of data, # of processors / ram / computations done per interaction
No. Those are ambiguous methods of determining a models performance. What you described indicates computational/data efficiency. Absolutely none of those metrics indicate the performance of a models output. To do that, one needs 1) a reference dataset and 2) a loss/reward metric or some ensamble of them. Computational efficiency has zero correlation to model performance.
I do this for a living.
> I assume
In this case, you assume incorrectly.
> pure computational power, measurable access to amount of data, # of processors / ram / computations done per interaction
No. Those are ambiguous methods of determining a models performance. What you described indicates computational/data efficiency. Absolutely none of those metrics indicate the performance of a models output. To do that, one needs 1) a reference dataset and 2) a loss/reward metric or some ensamble of them. Computational efficiency has zero correlation to model performance.
I do this for a living.
(post is archived)