Aaron Tay
@aarontay.bsky.social
I'm librarian + blogger from Singapore Management University. Social media, bibliometrics, analytics, academic discovery tech.
Pinned
Aaron Tay
@aarontay.bsky.social
· Nov 28
I'm an academic librarian blogger at Musings about librarianship since 2009.
To get a taste of what I blog, see "Best of..." Musings about librarianship - Posts on discovery, open access, bibliometrics, social media & more. Have you read them?
musingsaboutlibrarianship.blogspot.com/p/best-of.html
To get a taste of what I blog, see "Best of..." Musings about librarianship - Posts on discovery, open access, bibliometrics, social media & more. Have you read them?
musingsaboutlibrarianship.blogspot.com/p/best-of.html
[Watched] Finally watched this CNI talk from end to end. This pretty much sums up the cutting-edge issues regarding library + ai discovery. Library nexus, library workflow, library MCP servers. Also excellent q&a discussion
youtu.be/SgFsxlCQvOE?...
youtu.be/SgFsxlCQvOE?...
Discovery and Use Reimagined: Connecting Scholarly Collections and Artificial Intelligence Workflows
YouTube video by CNI: Coalition for Networked Information
youtu.be
February 18, 2026 at 7:50 PM
[Watched] Finally watched this CNI talk from end to end. This pretty much sums up the cutting-edge issues regarding library + ai discovery. Library nexus, library workflow, library MCP servers. Also excellent q&a discussion
youtu.be/SgFsxlCQvOE?...
youtu.be/SgFsxlCQvOE?...
[Watched] youtu.be/oXqJQeHWhUs?... - good refresher on doing systematic review with undermind.ai, 2dsearch, polygon - I must say many parts even free version of Undermind was heavily praised . Eg the email alert, lack of hallucinations, positive reviews by librarians etc. Not surprising I guess
YouTube
Share your videos with friends, family, and the world
youtu.be
February 18, 2026 at 6:33 PM
[Watched] youtu.be/oXqJQeHWhUs?... - good refresher on doing systematic review with undermind.ai, 2dsearch, polygon - I must say many parts even free version of Undermind was heavily praised . Eg the email alert, lack of hallucinations, positive reviews by librarians etc. Not surprising I guess
Even librarian have half a dozen definitions of what library databases or search engine mean. Eg i saw a systematic review librarians suggest database is say medline but platform/interface eg ovid, pubmed, Ebscohost, ProQuest, web of science is "search engine"
Typically librarians calling Google or Google Scholar a search engine & Scopus a database are gesturing towards a definition of "database" as closed, bounded, licensed. This isnt the distinction used in CS (3)
February 18, 2026 at 12:39 PM
Even librarian have half a dozen definitions of what library databases or search engine mean. Eg i saw a systematic review librarians suggest database is say medline but platform/interface eg ovid, pubmed, Ebscohost, ProQuest, web of science is "search engine"
The general sense I get is Claude currently is best at searching then Chatgpt with Gemini the worst at least for the non deep research versions.
February 18, 2026 at 10:39 AM
The general sense I get is Claude currently is best at searching then Chatgpt with Gemini the worst at least for the non deep research versions.
"On our API, Claude’s web search and fetch tools now automatically write and execute code to filter and process search results, keeping only relevant content in context—improving both response quality and token efficiency." This is going to be very powerful for academic search engines using claude
The Model You’ve All Been Waiting For:
Sonnet 4.6
SOTA performance in office tasks and financial reasoning 👈 (you’re seeing this too, right?)
www.anthropic.com/news/claude-...
Sonnet 4.6
SOTA performance in office tasks and financial reasoning 👈 (you’re seeing this too, right?)
www.anthropic.com/news/claude-...
February 18, 2026 at 10:28 AM
"On our API, Claude’s web search and fetch tools now automatically write and execute code to filter and process search results, keeping only relevant content in context—improving both response quality and token efficiency." This is going to be very powerful for academic search engines using claude
Curious about papers that investigated how good openalex was when used solely as database for systematic review. Undermind as usual found the top 2 super relevant, 1 that is closely relevant. Claude research mode found 2, OpenAlex deep research 1, Gemini Deep research 1, Consensus Deep search 0
February 17, 2026 at 12:39 PM
Curious about papers that investigated how good openalex was when used solely as database for systematic review. Undermind as usual found the top 2 super relevant, 1 that is closely relevant. Claude research mode found 2, OpenAlex deep research 1, Gemini Deep research 1, Consensus Deep search 0
[Watched] Using AI for systematic review: 2025 vs 2026
- interesting point PRISMA 2020 is challenged or constrained by new AI tools. Step by step approach not appropriate for newer methods eg AI can extract PICO from title/abstact, is that screening or "data extraction" ?
youtu.be/3ctb-Icri0o?...
- interesting point PRISMA 2020 is challenged or constrained by new AI tools. Step by step approach not appropriate for newer methods eg AI can extract PICO from title/abstact, is that screening or "data extraction" ?
youtu.be/3ctb-Icri0o?...
Using AI for Systematic Reviews: 2025 vs 2026 [A 15-Minute Wrap-Up]
YouTube video by Farhad Shokraneh
youtu.be
February 17, 2026 at 11:25 AM
[Watched] Using AI for systematic review: 2025 vs 2026
- interesting point PRISMA 2020 is challenged or constrained by new AI tools. Step by step approach not appropriate for newer methods eg AI can extract PICO from title/abstact, is that screening or "data extraction" ?
youtu.be/3ctb-Icri0o?...
- interesting point PRISMA 2020 is challenged or constrained by new AI tools. Step by step approach not appropriate for newer methods eg AI can extract PICO from title/abstact, is that screening or "data extraction" ?
youtu.be/3ctb-Icri0o?...
Explains a lot
February 17, 2026 at 10:17 AM
Explains a lot
Reposted by Aaron Tay
"How do we conceptualise equity, liberty, what counts as progress, what kind of economic growth, and whose knowledge counts as legitimate knowledge? These are not abstract. They are woven into the system. Into funding, journals, rankings, intellectual life, and scientific methods themselves."
Metascience for whom? A question as old as science.
Before we fix science, we need to ask who built it!
medium.com
February 17, 2026 at 8:03 AM
"How do we conceptualise equity, liberty, what counts as progress, what kind of economic growth, and whose knowledge counts as legitimate knowledge? These are not abstract. They are woven into the system. Into funding, journals, rankings, intellectual life, and scientific methods themselves."
Studied a bit more on how Claude skill works support.claude.com/en/articles/... - think we will be seeing a lot of librarians learning this soon
How to create custom Skills | Claude Help Center
support.claude.com
February 17, 2026 at 3:30 AM
Studied a bit more on how Claude skill works support.claude.com/en/articles/... - think we will be seeing a lot of librarians learning this soon
[Blogged] Are AI Tools Killing Review Articles? Two Failure Modes Suggest Otherwise aarontay.substack.com/p/are-ai-too...
Are AI Tools Killing Review Articles? Two Failure Modes Suggest Otherwise
arXiv recently restricted review article submissions in computer science, requiring journal or conference acceptance before deposit. They noted specifically that the change was driven by an “unmanagea...
aarontay.substack.com
February 16, 2026 at 4:54 PM
[Blogged] Are AI Tools Killing Review Articles? Two Failure Modes Suggest Otherwise aarontay.substack.com/p/are-ai-too...
Reposted by Aaron Tay
For example, one exercise I use is connecting to their own Zotero library through MCP and building different interfaces to explore it and find what works for their goals. Then it's easy to build a reusable skill for their preferred workflow. (2/2)
February 15, 2026 at 1:48 PM
For example, one exercise I use is connecting to their own Zotero library through MCP and building different interfaces to explore it and find what works for their goals. Then it's easy to build a reusable skill for their preferred workflow. (2/2)
The only thing close to this is SciSpace agents - which has many academic specific tools connected & it reasons what to do/use but its extremely expensive credit wise & probably not worth it. The DIY approach with setting up MCP servers + Claude Skills is still quite tricky currently(1)
the main diff is what I blogged about aarontay.substack.com/p/how-agenti... ie Consenus & its peers are running fixed flows, you can't suddenly ask arbitary tasks, e.g look at openalex, find all papers citing papers you wrote from 2023- and list are citing papers. You can with claude + openalex mcp
Deep Research, Shallow Agency: What Academic Deep Research Can and Can't Do"
The Agentic Illusion: Most Academic Deep Research runs fixed workflows and stumble when given unfamiliar literature review tasks that do not fit them.
aarontay.substack.com
February 16, 2026 at 5:33 AM
The only thing close to this is SciSpace agents - which has many academic specific tools connected & it reasons what to do/use but its extremely expensive credit wise & probably not worth it. The DIY approach with setting up MCP servers + Claude Skills is still quite tricky currently(1)
the main diff is what I blogged about aarontay.substack.com/p/how-agenti... ie Consenus & its peers are running fixed flows, you can't suddenly ask arbitary tasks, e.g look at openalex, find all papers citing papers you wrote from 2023- and list are citing papers. You can with claude + openalex mcp
Deep Research, Shallow Agency: What Academic Deep Research Can and Can't Do"
The Agentic Illusion: Most Academic Deep Research runs fixed workflows and stumble when given unfamiliar literature review tasks that do not fit them.
aarontay.substack.com
February 15, 2026 at 8:07 PM
the main diff is what I blogged about aarontay.substack.com/p/how-agenti... ie Consenus & its peers are running fixed flows, you can't suddenly ask arbitary tasks, e.g look at openalex, find all papers citing papers you wrote from 2023- and list are citing papers. You can with claude + openalex mcp
Spent time installing and playing Claude desktop + MCP servers like OpenAlex, semantic scholar, Zotero and I am starting to see why it might be worth it. Off the shelf AI literature review systems like Consensus can do all of this, including connect to Zotero & covers roughly the same index but(1)
February 15, 2026 at 8:03 PM
Spent time installing and playing Claude desktop + MCP servers like OpenAlex, semantic scholar, Zotero and I am starting to see why it might be worth it. Off the shelf AI literature review systems like Consensus can do all of this, including connect to Zotero & covers roughly the same index but(1)
Being blogging as a hobby for 17 years and the thing I notice is, the blogging content that does the best (most reads) tends to be those that are in my view more superficial, but easier to understand because they either repeat what is mainstream view or at best a small conceptual jump from it
February 14, 2026 at 3:04 AM
Being blogging as a hobby for 17 years and the thing I notice is, the blogging content that does the best (most reads) tends to be those that are in my view more superficial, but easier to understand because they either repeat what is mainstream view or at best a small conceptual jump from it
A better argument i think is around base rate of error. Around time of gpt3.5, the average error rate of answer was notably higher than textbooks or even reasonably knowledgeable person in area. That's why it was a bad idea to rely on it. But now? In many areas it is as good if not better
I know people will go but Aaron if you not an expert how do you know when the LLM is wrong? Thats a weird argument , replace the llm with a person or source you use to learn - dont you have the same problem? All ways of learning will have some chance of error which you as a novice can't detect (2)
February 14, 2026 at 2:07 AM
A better argument i think is around base rate of error. Around time of gpt3.5, the average error rate of answer was notably higher than textbooks or even reasonably knowledgeable person in area. That's why it was a bad idea to rely on it. But now? In many areas it is as good if not better
Frontier LLMs are really a huge force multlper if you love to learn and/or don't have much learning anxiety (1)
February 14, 2026 at 1:54 AM
Frontier LLMs are really a huge force multlper if you love to learn and/or don't have much learning anxiety (1)
In the early 2000s if you had the habit to use Google to answer questions vs someone who refused to do web search at all because you couldn't trust the sources, the former had a clear competitive advantage despite answer sometimes from Google vbeing wrong. (1)
February 13, 2026 at 11:48 AM
In the early 2000s if you had the habit to use Google to answer questions vs someone who refused to do web search at all because you couldn't trust the sources, the former had a clear competitive advantage despite answer sometimes from Google vbeing wrong. (1)
Elicit just announced they switched to the latest Opus 4.6 with uplift of 12% recall. Scite assistant gives you choice of models latest is Opus 4.5 & GPT5.2 with effort up to high. Meanwhile most legacy databases claim to be using the totally out of date 4o-mini from 2024 (which is retired today)(1)
February 13, 2026 at 6:49 AM
Elicit just announced they switched to the latest Opus 4.6 with uplift of 12% recall. Scite assistant gives you choice of models latest is Opus 4.5 & GPT5.2 with effort up to high. Meanwhile most legacy databases claim to be using the totally out of date 4o-mini from 2024 (which is retired today)(1)
Reading an email from a reader of my blog who thanks me but describes feeling learning anxiety even "existential dread" as they feel there's so much to learn and they are falling behind. Thing is, are you surprised to know I feel that way too sometimes (latest was around Claude code + skills/mcp(1)
February 13, 2026 at 2:36 AM
Reading an email from a reader of my blog who thanks me but describes feeling learning anxiety even "existential dread" as they feel there's so much to learn and they are falling behind. Thing is, are you surprised to know I feel that way too sometimes (latest was around Claude code + skills/mcp(1)
Spent more time playing with tools that look up openalex, crossref etc to check for hallucinations in references. Tried scispace and consensus generated reference lists. Basically out of say 40, the tool would flag 1/2 as potential hallucinations, but on checking the tool is almost always wrong. (1)
February 12, 2026 at 6:06 PM
Spent more time playing with tools that look up openalex, crossref etc to check for hallucinations in references. Tried scispace and consensus generated reference lists. Basically out of say 40, the tool would flag 1/2 as potential hallucinations, but on checking the tool is almost always wrong. (1)
I know people do cite my blog posts and I dont really need to play the publish or perish game but still to get proper credit for my intellectual ideas (not many that are really novel & valuable but there are some), I will have to publish something somewhat formal so people have something to cite.
February 12, 2026 at 12:15 AM
I know people do cite my blog posts and I dont really need to play the publish or perish game but still to get proper credit for my intellectual ideas (not many that are really novel & valuable but there are some), I will have to publish something somewhat formal so people have something to cite.
Life is great when you have so many exciting things to do and you are not sure which to do first...
February 11, 2026 at 11:34 PM
Life is great when you have so many exciting things to do and you are not sure which to do first...
By now librarians should learn that arguments like tool x cannot perfectly do y hence it can never be useful are just rubbish. You need actually evaluate it holistically & not dismiss it with a argument like this. But we never learn and keep doing it..web search, Google Scholar, Wikipedia, LLMs ...
February 11, 2026 at 1:25 PM
By now librarians should learn that arguments like tool x cannot perfectly do y hence it can never be useful are just rubbish. You need actually evaluate it holistically & not dismiss it with a argument like this. But we never learn and keep doing it..web search, Google Scholar, Wikipedia, LLMs ...