Aaron Tay
@aarontay.bsky.social
I'm librarian + blogger from Singapore Management University. Social media, bibliometrics, analytics, academic discovery tech.
Quick test that the Pubmed MCP/Connector works like normal Pubmed. Query below and shows top 5 (can be changed to top X) by PMID and total number of results. Tested in Pubmed advanced mode, same number of results, and top 5 PMID match by most recent (4)
November 25, 2025 at 4:50 AM
Quick test that the Pubmed MCP/Connector works like normal Pubmed. Query below and shows top 5 (can be changed to top X) by PMID and total number of results. Tested in Pubmed advanced mode, same number of results, and top 5 PMID match by most recent (4)
Have to check if the pubmed Mcp server actually works correctly for long Boolean strings but I've always thought human Vs GPT competitions on Boolean strings crafting was always a bit unfair cos humans could test their strategies. This evens things out. Maybe add a mesh browser access via MCP? (2)
November 25, 2025 at 12:49 AM
Have to check if the pubmed Mcp server actually works correctly for long Boolean strings but I've always thought human Vs GPT competitions on Boolean strings crafting was always a bit unfair cos humans could test their strategies. This evens things out. Maybe add a mesh browser access via MCP? (2)
Was testing out how good Gemini 3 pro, GPT5.1, Opus 4.5 was at crafting Boolean for systematic review in pubmed & asking to compare & I suddenly realised because Claude has access to Pubmed MCP it actually uses that to TEST and evaluate search strings just like humans! (1)
November 25, 2025 at 12:45 AM
Was testing out how good Gemini 3 pro, GPT5.1, Opus 4.5 was at crafting Boolean for systematic review in pubmed & asking to compare & I suddenly realised because Claude has access to Pubmed MCP it actually uses that to TEST and evaluate search strings just like humans! (1)
Assuming it searches ONCE (it seems to use multiple query combine them) then goes down the results using Gemini to evaluate relevancy the fact it stops after looking at the first 300 is similar to the rule some evidence synthesis people use for GS to stop after 200-300!
November 19, 2025 at 11:14 AM
Assuming it searches ONCE (it seems to use multiple query combine them) then goes down the results using Gemini to evaluate relevancy the fact it stops after looking at the first 300 is similar to the rule some evidence synthesis people use for GS to stop after 200-300!
Hmm seems to be stopping after evaluating top 300?
November 19, 2025 at 11:03 AM
Hmm seems to be stopping after evaluating top 300?
Another semantic equalvant version and it also stopped at 50 relevant results after looking at top 90 or so results.. so likely it is coded to stop when 50 relevant results are found. BTW There are some false positives in what it considers relevant but minor as it is easy to see and exclude
November 19, 2025 at 10:54 AM
Another semantic equalvant version and it also stopped at 50 relevant results after looking at top 90 or so results.. so likely it is coded to stop when 50 relevant results are found. BTW There are some false positives in what it considers relevant but minor as it is easy to see and exclude
Very interesting. A query with many more hits, allowed more iterations but eventually stopped at 50 relevant results and only looked at top 100 or so results. I suspect the 50 is too round a number to be by chance
November 19, 2025 at 10:48 AM
Very interesting. A query with many more hits, allowed more iterations but eventually stopped at 50 relevant results and only looked at top 100 or so results. I suspect the 50 is too round a number to be by chance
another run. stopped at around 300 results..or does it always stop around there....
November 19, 2025 at 10:43 AM
another run. stopped at around 300 results..or does it always stop around there....
Finally stopped at 31 found relevant results with no way to search further, Fascinating, it has some stopping point/rule. This is giving me some undermind type vibes...
November 19, 2025 at 10:35 AM
Finally stopped at 31 found relevant results with no way to search further, Fascinating, it has some stopping point/rule. This is giving me some undermind type vibes...
After the 2nd search it searched almost 200 results and claims to find 20 relevant results. Let me go further and cick more results (5)
November 19, 2025 at 10:32 AM
After the 2nd search it searched almost 200 results and claims to find 20 relevant results. Let me go further and cick more results (5)
Fascinating the side bar shows how many papers GS has evaluated . It eventually stopped going up and displayed 10 relevant results found.. fascinating..
(4)
(4)
November 19, 2025 at 10:28 AM
Fascinating the side bar shows how many papers GS has evaluated . It eventually stopped going up and displayed 10 relevant results found.. fascinating..
(4)
(4)
Every 10 results you get a "more results" button (3)
November 19, 2025 at 10:22 AM
Every 10 results you get a "more results" button (3)
Runs for around 10-15s and gets some results. Typically around 5 results with generated text on why it matches. As you scroll down it seems to be either generating text to "explain" already ranked results or it is doing work to pull in more top results. (2)
November 19, 2025 at 10:22 AM
Runs for around 10-15s and gets some results. Typically around 5 results with generated text on why it matches. As you scroll down it seems to be either generating text to "explain" already ranked results or it is doing work to pull in more top results. (2)
an obvious thng to do and the paper acknowledges is to do field nornalization. Obviously non hard/stem areas are disadvantaged. (11)
November 18, 2025 at 7:52 AM
an obvious thng to do and the paper acknowledges is to do field nornalization. Obviously non hard/stem areas are disadvantaged. (11)
As such this is their current Scite index (SI). It is essentially squaring the USI (unweighted scite index) then multiply by total (normal) citations and then do a natural log. Essentially, we are using the normal citation count as a base, then adjusting based on the unweight scite index (7)
November 18, 2025 at 7:42 AM
As such this is their current Scite index (SI). It is essentially squaring the USI (unweighted scite index) then multiply by total (normal) citations and then do a natural log. Essentially, we are using the normal citation count as a base, then adjusting based on the unweight scite index (7)
Also it seems USI favours hard science, so if you use it by institution, journal, field, it is obvious hard science in particular very specific technical institutions, journals etc have an advantage and not the usual institutions or glamor journals (5)
November 18, 2025 at 7:38 AM
Also it seems USI favours hard science, so if you use it by institution, journal, field, it is obvious hard science in particular very specific technical institutions, journals etc have an advantage and not the usual institutions or glamor journals (5)
But leaving that aside Unweighted Scite index has a few interesting properties. Firstly at the journal level USI seems uncorrelated to JIF. Then again the range of USI is relatively small from 0.8-0.9 ? (4)
November 18, 2025 at 7:34 AM
But leaving that aside Unweighted Scite index has a few interesting properties. Firstly at the journal level USI seems uncorrelated to JIF. Then again the range of USI is relatively small from 0.8-0.9 ? (4)
First they define a USI or unweighted scite index which is a simple ratio of supported citation over (supported + contrasting citations). I assume here you are familar with scite citing types. But if not see direct.mit.edu/qss/article/... for technical paper on how they use ML to train (2)
November 18, 2025 at 7:27 AM
First they define a USI or unweighted scite index which is a simple ratio of supported citation over (supported + contrasting citations). I assume here you are familar with scite citing types. But if not see direct.mit.edu/qss/article/... for technical paper on how they use ML to train (2)
When importing I notice not all of the 30 sources can be succesfuuly imported. A few were highlighted in red, and when you mouse over the icon you are told they are either behind paywall or invalid URL? (4)
November 14, 2025 at 10:36 AM
When importing I notice not all of the 30 sources can be succesfuuly imported. A few were highlighted in red, and when you mouse over the icon you are told they are either behind paywall or invalid URL? (4)
It definitely seems to execute much faster than normal Gemini Deep research finding only 30 sources. But to be fair, this mode is more of "Deep Search" cos it will only find the top sources but does not synthese the results. From there you can import the sources into NotebookLLM. (3)
November 14, 2025 at 10:34 AM
It definitely seems to execute much faster than normal Gemini Deep research finding only 30 sources. But to be fair, this mode is more of "Deep Search" cos it will only find the top sources but does not synthese the results. From there you can import the sources into NotebookLLM. (3)
1. Go to discover sources then choose between Fast Research and Deep Research. You can also choose between "web" and files in "google drive". Compared to Gemini Deep Research, it does not show you it's plan and goes off immediately (20
November 14, 2025 at 10:26 AM
1. Go to discover sources then choose between Fast Research and Deep Research. You can also choose between "web" and files in "google drive". Compared to Gemini Deep Research, it does not show you it's plan and goes off immediately (20
interesting read on business models.
November 13, 2025 at 2:57 AM
interesting read on business models.
Interesting new Google Scholar PDF button feature scholar.googleblog.com/2025/11/mark...
November 12, 2025 at 6:06 AM
Interesting new Google Scholar PDF button feature scholar.googleblog.com/2025/11/mark...
tried again with "research mode" so it thinks harder.
November 8, 2025 at 7:40 PM
tried again with "research mode" so it thinks harder.
Seems to do okay getting to the figure or table (3)
November 8, 2025 at 7:27 PM
Seems to do okay getting to the figure or table (3)