Does it mean one shouldn't use specialised academic ai search tools like Elicit, Undermind, Consensus? More testing needed but the gap is definitely diminishing if not gone if all you care is retrieval capabilities but affordances & other features provided by specialised tools are still worth it

Undermind is the only specialised academic search I tested that put up a good fight finding 2-3 relevant papers at the top app.undermind.ai/report/a591d.... But it still fails to find grey literature like www.fhi.no/en/publ/2023... that Claude spotted.

As you can see below Claude easily spots the critical Stansfield et al. (2025) paper by searching Consensus. Curious that Consensus Deep Search that does a ton more searches either fails to find it or somehow ranks it so low it isnt picked up?

Oddly, I notice now Claude has a connector for Consensus and when I ask Claude Sonnet 4.6 (web) to use Consesnus to search for the same question, the results are way better. claude.ai/share/362771... - it probably has nothing to do with the source but the way Consensus (web) "chooses" to search

This is Consensus Deep Search that totally failed to find any relevant paper that tested Openalex for SR vs golden standard papers consensus.app/search/can-y...

Here's using the latest Claude Sonnet 4.6 + research mode. claude.ai/public/artif... no special connectors.

Personally these days i would recommend undermind + Claude/openai/gemini deep research to cover more ground. The general DR are mostly
Irriating to use due to layout issues but their retrieval capabilities are truly impressive

In fact, I think Claude Research mode won, mostly because it could find grey literature stuff by EPPI centre that Undermind couldnt see. But that's the general pattern im seeing, in terms of pure retrieval the generic Deep research are matching the academic ones even if you exclude grey literature.

Like Google Scholar, likely 99% of relevant papers are indexed but can you actually retrieve it? This study looks interesting pmc.ncbi.nlm.nih.gov/articles/PMC...

Also talks about how PRISMA assumes dedupe as a single stage but AI enables continuous deduping and searching throughout the whole process (3)

Claims also that searching a single database like openalex can reliably retrieve most relevant papers which challenges PRISMA 2020 multi database assumption. Is this generally believed? That you can search openalex only and get similar practical recall vs multi database? (2)

Not really. Maybe open.substack.com/pub/aarontay... . Same thing happening when papers with ghost references are surfaced on platforms with library link resolvers and clicking gets you a dynamically generated record which you can request via ILL in library catalogue

Code execution with MCP: building more efficient AI agents \ Anthropic www.anthropic.com/engineering/...

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)