[torginus]

In my experience LLMs suck at (product) recommendations - I was looking for books with certain themes, asked ChatGPT 5, the answer was vague, generic and didn't fit the bill. At another time I writing an essay and was looking for famous figures to cite as examples of an archetype, and ChatGPT's answers were barely related.

In both cases, LLMs gave me examples that were generally famous, but very tangentially related to the subject at hand (at times, ChatGPT was reaching or straight up made up stuff).

I don't know why it has this bias, but it certainly does.

[Anon1096]

I work on rec systems.

The ideal here will be a multi tiered approach where the LLM first identifies that a book should be recommended, a traditional recommendation system chooses the best book for the user (from a bank of books that are part of an ads campaign), and then finally the LLM weaving that into the final response by prompt suggestion. All of this is individually well tested for efficacy within the social media industry.

I'll probably get comments calling this dystopian but I'm just addressing the claim that LLMs don't do good recommendations right now, which is not fundamental to the chatbot system.

[torginus]

All this would imply that the core value derives from better rec systems and not LLMs, which will merely embed the recommendation into their polite fluff.

Rec systems are in use right now everywhere, and they're not exactly mindblowing in practice. If we take my example of books with certain plotlines, it would need some super-high quality feature extraction from books (which would be even more valuable imo, than having better algorithms working on worse data). LLMs can certainly help with that, but that's just one domain.

And that would be a bespoke solution for just books, which would, if worked, would work with a standard search bar, no LLM needed in the final product.

We would need people to solve every domain for recommendation, whereas a group of knowledgeable humans can give you great tips on every domain they're familiar with on what to read, watch, buy to fix your leaky roof, etc.

So in essence, what you suggest would amount to giving up on LLMs (except as helpers for data curation and feature extraction) and going back to things we know work.