Talking about vibe coding @openalex claims that the GUI is dead and shows how a simple prompt in Claude (not even Claude code!) can create nice analysis and even web app! www.youtube.com/watch?v=q9KF.... Is it that easy? I gave it a try (1)

Vibe coding is a super power for instruction. I always struggled to explain cosine similarity in vector search in workshops for librarians. Today I saw a participant comment she finally understood what cosine has to do with vector search. Maybe its coincidence but this time I let (1)

I find it sad that people are still wondering if "AI search" is valuable for academic search. The answer is clearly yes, but i can see why it's sometimes hard to see because imho many product on the market are barely move the needle (and some early versions are worse than status quo!) but (1)

Finished Day 1 of our AI-powered Search in libraries workshop. Day 2 on Wednesday is when it gets intense :)

[Read] Ten years of a ‘quiet culture war’: where does it stand now? insights.uksg.org/articles/10..... Nothing really new it seems

Point in question. I spent a couple of hours trying to write a post to explain a family of techniques called Learning to Rank (LTR) a technique that was probably sota in the mid 2010s before more neural methods, particularly transformer based methods eg BERT took over as rerankers but... (1)

It is okay for librarians to be skeptical of new technology. Nay, you SHOULD. Some Information professionals at the dawn of Google Scholar really scrutinized it, exposing a lot of issues in terms of coverage & metadata cleanliness issues. Today GS is heavily used doesnt mean they were wrong (1)

I was recently working on a trial/pilot with a vendor on a new tool/service. He kept advising me to use as a "success factor", % of active users out of those that trialled. While I understand the advantages of doing focused marketing, it seems odd to (1)

I've been a librarian almost 20 years, a people manager and "head" for a much shorter period but when I reflect on what works, I have actually quite strong and possibly even controversial views in terms of what successful library leadership and management look like (1)

Primo really goes into detail about the relevancy ranking behind the hood but I doubt even one in a hundred librarians have really looked at it. Even few fully understand it. knowledge.exlibrisgroup.com/Primo/Conten...

Started accepting tips on Substack, partly as an experiment. Wasn't sure if anyone would actually donate. Gratified that some readers have chosen to support the work. My blog (Musings About Librarianship) is 100% independent & not sponsored If you find value in the content: ko-fi.com/aarontay

[Blogged] The Blank Box Problem: Why It's Harder Than Ever to Know What to Type Into an AI Search Bar - open.substack.com/pub/aarontay... . A bit over written but has some interesting points i think

It is 2001, to discourage students from using Google, I'm going to call them "Automated Query-String Matching Engine with Advertising-Optimised Relevance Ranking", i am sure that clever name will shame them from using it

I'm slow. I only realised after 6 months on substack that you could easily create footnotes, before that I was basically using callouts/sidebars. But both have their uses I guess.

Pretty scary how fast things move. Looking at slides I created in May 2025 for a workshop on retrieval and AI, it's still mostly relevant—but whether due to my deeper understanding or actual advances (e.g. Rise of MCP), parts of it look oversimplified now (e.g RAG vs reasoning models+search).

Most of my last 2–3 years can be summed up by this post. The belief that information retrieval is going through a fundamental shakeup, librarians really need to study the nuts & bolts of IR to be helpful guides to our users. But sometimes I think I might be leading people on a quixotic quest (1)

Data tables feature in Google NotebookLM is nice, too bad you cannot click on the values to see what part of the source it relied on to extract the data.

Personas of librarians and their response to AI - part 2. This is all tongue in cheek. substack.com/@aarontay/no...

[Watched] eCAUG 2025 Day 2: Primo Research Assistant versus Other Academic AI search tools youtu.be/Mi-Bo1wRlFA?... (1)

Semantic Search Without Embeddings softwaredoug.com/blog/2026/01... - insightful

Bad argument incoming - If genai tools were as good as they say, they'd be easy to use & their uses obvious without training. But they're not and so instead they just constantly tell us to use these tools + downplay the cons. Spot the massive flaw?

Dismissing some technology as 'not AI' or giving it names to shame it isn't going to alter one whit how useful it is. It's a convenient and lazy way to avoid actually trying it, of course to see if it's really useful (1)

Feedback on my post has been very good. But I wonder how many read to the appendix. I think the framework here is probably the most accurate but what vendors will not be that transparent & many librarians would not have enough info retrieval chops currently to understand the differences

Playing with the new natural language search in Primo NDE. Fascinating the query interpreting feature where it converts natural language into Boolean is more capable than Primo RA. Eg it recognises more filters! Eg find me <resource type> on x mostly works. Primo RA you only allowed 3 values

anyone who has or knows a library that has gone live with Primo NDE?

[Blogged] What do we mean by AI-powered Search? open.substack.com/pub/aarontay...

[Read] 2025 AI roundups open.substack.com/pub/simonw/p... vs open.substack.com/pub/sebastia... - mostly the same points

I first wrote abt perplexity in late 2022 but my use fell off as general LLMs like chatgpt, Claude, Gemini became great at adding search as a tool (o3 particularly was huge jump in capabilities). Trying it out again with perplexity pro.... its not bad but somehow the hallucination rate seems high

This year on Bluesky I wrote 569 posts and 1,578 replies. I received 4,440 likes, whereas 68 was from my most popular post, and apparently I love saying "search" and ✅! www.madebyolof.com/bluesky-wrap...

One hour to 2026 here. This year I shifted my blog to substack and got a second wind hitting 32 posts! My highest output in recent years only exceeded in the first few years of my blog in 2009,2010 etc! Hope to keep this up! For readers of my blog, thanks for the support and I will see you in 2026.

[Last blog post of 2025] "AI-powered search" hides at least 4 different things—post-retrieval features, semantic search, LLM ranking, and synthesis. Your concerns about one may not apply to another. New post unpacking what's actually under the hood. aarontay.substack.com/p/what-do-we...

Belatedly learning about the great split in the wikidata query service

How do you know you have acquired a bit of understanding of a technical area? When you write something & you immediately can see that what you wrote is a bit of a simplification but it would take a lot more effort to address the technical point and it's not worth it usually .

I must say im very impressed with Claude Opus 4.5. Its really seems smarter than Gemini 3.5 pro and chatgpt5.2+thinking (20 bucks tier). Part of it is the memory, chatgpt has that too but Claude is more specific. Beyond that Claude hedges a lot less than the other two when asked to critique

So much Performative BS on linkedin. Eg someone not actually reading the paper and linking to it with a point that doesn't make sense. Typical performative "engagement" with preprints based on title keywords rather than substance.

I find it funny, that the anti-ai camp people often sound like stochastic parrots themselves, eg they will link to papers that are totally irrelevant (or the link is so tenuous if you are generous) to the point they are making , which you do see happen sometimes with RAG systems.

Years ago a librarian (who was asked this question by a prof) asked me what's the difference between a database and search engine and even then I knew it was a tricky question. Part of it is what CS calls "database" or "search engine/index" is different from the LIS definition. (1)

[blogged] Why Ghost References Still Haunt Us in 2025—And Why It's Not Just About LLMs aarontay.substack.com/p/why-ghost-...

I have a confession. Ever since the middle of this year? I realised the frontier LLMs, eg Claude Sonnet 4.5 became good enough to expand my ideas (eg from a social media thread) to a full blown post, I've been extraordinarily productive - im currently sitting on 3-5 almost completed posts...(1)

[Read] Performance of AI tools in citing retracted literature preprints.jmir.org/preprint/88766 interesting paper but similar to others i have read but I'm willing to bet this issue will likely disappear or be so reduced to point not worth worrying in 2 years max. Why? (1)

Interesting… That affiliation data could be rebuilt with ORCID if that data is still getting through to Crossref. @aarontay.bsky.social wrote about the disappearance of Elsevier abstracts from OpenAlex earlier this year aarontay.substack.com/p/the-petrol...

A fallacy in my thinking. I assumed RAG properly done will reduce ghost ref to near zero. But this is not true if your source you do retrieval over can get contaminated which is easy if you just search the web! Elicit, Consensus searching indexes like Semantic Scholar should be less vulnerable

Really confused by all these reports from librarians & bloggers about them getting a ton of fake citations (in the sense of paper doesnt exist at all). While I buy this being common in 2023, by 2025 this shouldn't be that common given even free versions can search the web.

www.straitstimes.com/life/arts/th... Guess someone not letting this go...

More thoughts about using LLMs to generate Boolean either manually or automated in system (eg Primo RA, Ebsco natural language search) why some librarians find it useful & some don't. Spoiler It has to do with use case & expectations (1)

Preparing for perhaps my last talk of 2025

[Blogged] Deep Research, Shallow Agency: What Academic Deep Research Can and Can't Do open.substack.com/pub/aarontay...

I really dont understand why whenever I see someone post a big fail using LLMs online. But when I try it I can't reproduce it.. Sometimes they are using a free weaker model but that happens enough it can't be just that. Or maybe im just lucky?

Main point here is boolean + minimal relevance ranking with bm25 typically gives you mediocre precision@5/10 & theres not much you can do to improve. "AI search", matching that or even slightly worse is actually not a good result. Modern retrieval methods can not only give better precision@10 (1)

Insight: I've been quite negative about those simple "push natural language input to LLM to generate Boolean & run" systems than many librarians mostly because I have higher expectations than them.. (1)