[torginus]

I mean I did data science and ML (more the engineering part) back when it was popular, and I kinda echo those sentiments.

ML folks told me their methods would only outperform traditional ones if we could feed the training pipelines an order of magnitude more data than we had. And if a process we're trying to predict only produces X amount of data over a meaningful timeframe, getting 10x, or 100x the amount usually requires 'lateral thinking', which even at the best of times meant either building mass-survelliance into the product or pulling in tangential data, to get marginally better results at best.

What actually worked was understanding what data we had, what it meant, what we missed, and after gathering that right information, even the simplest mathematical tools yielded results.

They still ended feeding the entire thing into 'fancy algorithm X' from scikit-learn, crunched the numbers on those fancy GPUs and got a 5% better result than what the simple methods gave us.

It was still framed as the triumph of these fancy algorithms and expensive data scientists and not people with domain expertise working hard to find and get the right data.

I'm not even salty at this point, I'm just kinda disappointed that a big org has such a hard time doing the right thing for the right reasons.