4 ms·

I’m curious as to what readers think Large Language Models will bring to the recommendations table that our current machine learning doesn’t.

trollbridge 2mo ago
I’m curious too, as ML is pretty amazing; LLMs are one application of ML; LLMs have a much narrower good set of uses than broad ML. You can use LLMs for things they are a bad fit for. I know someone who uses it like a spreadsheet to add sums of numbers, etc.
cjs_ac 2mo ago
It will bring the ability to tell investors that Netflix is using Large Language Models.
- dgellow 2mo ago
  Yeah, I think that’s the cynical and correct answer, they don’t have the explosive growth from last decade and need something to boost their share price
porridgeraisin 2mo ago
In this setting, they are essentially using it as a feature extractor. As for moving to this versus your bespoke feature extractor, _given_ that your existing features and their compositions are not antithetical to their language representation, an LLM will be an equal or better feature extractor. But for example, considering the basic feature genre, if your collaborative filter has learnt features that smartly recommending a show with one text feature "sci-fi" to people whose preferences have the text feature "comedy" because of learned behaviour despite the text, then you have to verbalise this feature "scifi,laugh track" or maybe providing samples of the subtitles of the show or add a "frequently co-watched with" section (which contains comedy shows) to the prompt, to effectively get an LLM to do the same thing (or many other ways to induce a hybrid embedding) In many cases, people have almost entirely verbalizable features and feature compositions in their existing systems even if it may not exactly be optimal. So it's a good idea to try out LLMs there. Composition mentioned everywhere above is crucial. Provided you can verbalise your important features, LLMs can perform very strong deductions and compositions out of the box above and beyond our own feature interactions that we use with say xgboost setups. And it's dynamic in the sense that it gives you a foundation model you don't have to retrain to use new (verbalizable) interactions.
watwut 2mo ago
Netflix recommendations are bad. So, I expect them to be bad differently.

trollbridge 2mo ago

I’m curious too, as ML is pretty amazing; LLMs are one application of ML; LLMs have a much narrower good set of uses than broad ML. You can use LLMs for things they are a bad fit for. I know someone who uses it like a spreadsheet to add sums of numbers, etc.

cjs_ac 2mo ago

It will bring the ability to tell investors that Netflix is using Large Language Models.

dgellow 2mo ago
Yeah, I think that’s the cynical and correct answer, they don’t have the explosive growth from last decade and need something to boost their share price

porridgeraisin 2mo ago

In this setting, they are essentially using it as a feature extractor. As for moving to this versus your bespoke feature extractor, _given_ that your existing features and their compositions are not antithetical to their language representation, an LLM will be an equal or better feature extractor. But for example, considering the basic feature genre, if your collaborative filter has learnt features that smartly recommending a show with one text feature "sci-fi" to people whose preferences have the text feature "comedy" because of learned behaviour despite the text, then you have to verbalise this feature "scifi,laugh track" or maybe providing samples of the subtitles of the show or add a "frequently co-watched with" section (which contains comedy shows) to the prompt, to effectively get an LLM to do the same thing (or many other ways to induce a hybrid embedding) In many cases, people have almost entirely verbalizable features and feature compositions in their existing systems even if it may not exactly be optimal. So it's a good idea to try out LLMs there. Composition mentioned everywhere above is crucial. Provided you can verbalise your important features, LLMs can perform very strong deductions and compositions out of the box above and beyond our own feature interactions that we use with say xgboost setups. And it's dynamic in the sense that it gives you a foundation model you don't have to retrain to use new (verbalizable) interactions.

watwut 2mo ago

Netflix recommendations are bad. So, I expect them to be bad differently.