One problem i think with a lot of librarian discourse on "Ai" is its all very high level. While some of this is valuable I think there is more value in looking at specific use cases. I'm focusing on "ai in search" and that itself is very broad and requires further breakdown (1) - ThreadSky

aarontay.bsky.social • 2 days ago

One problem i think with a lot of librarian discourse on "Ai" is its all very high level. While some of this is valuable I think there is more value in looking at specific use cases. I'm focusing on "ai in search" and that itself is very broad and requires further breakdown (1)

Comments

aarontay.bsky.social•2 days ago

Do we mean 1) use chatbot to return answers 2) RAG enabled search 3) "semantic search" that goes beyond keyword search? I would argue #1 is clearly objectionable, #2 is debatable #3 is relatively benign though we have to be aware of cons (2)

aarontay.bsky.social•2 days ago

People might be surprised to know pubmed has "ai", leaving aside Auto term mapping, auto indexing of MesH, the search algo involves the use of BM25 followed by a reranker step. State of art now probably involves some transformer based model. Nobody thinks this is a problem (3)

aarontay.bsky.social•2 days ago

But that's just reranking, the first step is still keyword based BM25. But now what if they use a LLM to translate to boolean search? Or if they out right use dense retrieval embeddings for first stage retrieval?
Then you might have diff tradeoffs on interpretability and reproducibility(4)

aarontay.bsky.social•2 days ago

There's lot of unknowns of course, are the embeddings biased in some unknown way beyond gender stereotypes etc. But i judge the risk to be not that high comparatively speaking to other uses of ai given unknown bias in google etc (5)

aarontay.bsky.social•2 days ago

Then there's the rise of deep search, agentic search. Where the LLM searches on your behalf. Even if all it does is pull up a list of relevant results like https://undermind.ai which is very popular here are there any long term effects of relying on it too much?(6)

aarontay.bsky.social•2 days ago

And then we move into use case 2, RAG enabled systems that not just give you potentially relevant items but try to extract answers whether generating a short summary or a full report (deep research). This is a far bigger change with big implications(7)

Posting Rules

Be respectful to others
No spam or self-promotion
Stay on topic
Follow Bluesky's terms of service

Comments

Posting Rules

Reply