I'm happy someone did this. People are making rumblings in traditional simulation modeling circles that ML models will replace traditional forward modelling...because I'm assuming a PM who hasn't touched code in a decade was impressed that chatGPT can write their emails for them.
I feel like it was never the case that dumb fitting could replace a first principles model you already have in hand, and I always had the concern that extrapolation with these models is just a prone to error as it is when your model is linear least squares for a scatterplot, you cannot fit to data that just isn't in your initial dataset! Sure you can initially but at some point your model breaks down as any model does, a priori, statistical, or whatever.
I could have always told you that (and I told my peers that for yeara) because the fact is intuitively obvious but it's great someone is actually publishing about it now. Now we just need similar papers for language models, which we sort of already experience when you ask chatGPT about 2023 events but somehow still cannot pierce the minds of the hypemen.
I feel like it was never the case that dumb fitting could replace a first principles model you already have in hand, and I always had the concern that extrapolation with these models is just a prone to error as it is when your model is linear least squares for a scatterplot, you cannot fit to data that just isn't in your initial dataset! Sure you can initially but at some point your model breaks down as any model does, a priori, statistical, or whatever.
I could have always told you that (and I told my peers that for yeara) because the fact is intuitively obvious but it's great someone is actually publishing about it now. Now we just need similar papers for language models, which we sort of already experience when you ask chatGPT about 2023 events but somehow still cannot pierce the minds of the hypemen.