

I agree with you on almost everything.
It’s like the opposite of classic ML, relatively tiny special purpose models trained for something critical, out of desperation, because it just can’t be done well conventionally.
Here i disagree. ML is using high dimensional statistics. There exist many problems, which are by their nature problems of high dimensional statistics.
If you have for an example an engineering problem, it can make sense to use an ML approach, to find patterns in the relationship between input conditions and output results. Based on this patterns you have an idea, where you need to focus in the physical theory for understanding and optimizing it.
Another example for “generative AI” i have seen is creating models of hearts. So by feeding it the MRI scans of hundreds of real hearts, millions of models for probable heart shapes can be created and the interaction with medical equipment can be studied on them. This isn’t a “desperate” approach. It is a smart approach.
The recognition of the pattern is done by the machine learning. That is the core concept of machine learning.
For the interpretation you need to use your domain knowledge. Machine learning together with knowledge in the domain analyzed can be a very powerful combination.
Another example in research i have heard about recently, is detection of brain tumors before they occur. MRIs are analyzed of people who later developed brain tumors to see if patterns can be detected in the people who developed the tumors that are absent in the people who didn’t develop tumors. This knowledge of a correlation between certain patterns and later tumor development could help specialists to further their understanding of how tumors develop as they can analyze these specific patterns.
What we see with ChatGPT and other LLMs is kind of doing the opposite by detaching the algorithm from any specific knowledge. Subsequently the algorithm can make predictions on anything and they are worth nothing.