My AI research assistant setup

As my days meander between rest, pain management, and gentle movements, there is some space for the restless brain to explore. This is where I have landed so far with my AI as research assistant set up.¹

Background

When I was undertaking my masters I interviewed a family friend who did his PhD in education directly after WWII had ended. It was 60 pages long, typewriter written, if you made a mistake on a page you had to retype the whole page. It sometimes took 3 months to get a book via interlibrary loan. I remember envying this slow thinking and writing.

Fast forward 21 years: when I explore any given topic there are thousands, if not tens of thousands of results. I don’t think the effects of publish or perish culture require referencing here.

The amount of time and effort it takes to identify and sort through these publications would exceed and deplete my resources. My time, my ADHD rabbit hole habit, and my self imposed expectation of thoroughness. Particularly, if I want to be at least aware of various voices and viewpoints a discourse holds.

Assistant 01

Google Scholar Labs and Google Scholar

I have begun to use Google Scholar Labs just before hospitalisation.

If you are a bilingual scholar this struggle will be familiar: I regularly was frustrated because I couldn’t find the right word or phrasing to search for what I was looking for.

Google Scholar Labs lets you put in an open text question, and being able to ask out loud is useful.

I then can use these results to dig deeper, find the right search terms. Sometimes, because the results give you a summary for each article or book, the results are enough to answer my question.

And as usual of course always open the resources and check for accuracy.

If you want a full review this blog post is well structured: https://effortlessacademic.com/google-scholar-labs-ai-feature-review-for-academics/

NotebookLM

My friend Tood had shown this to me last year, and I have been exploring the ever growing set of features.

The usefulness of this clicked for me, when one weekend I read 17 papers and then could not remember where this one quote came from.

I know, I know! Bad note making technique. So I chucked all the papers into NotebookLM and asked it to identify the source, couple of seconds later I got it.

Now I am using it to cull key points and arguments from papers. My process:

I search for papers and books, scan or skim them for usefulness and quality, and then add them into NBLM.
If it is something super interesting or seminal work I read deep but I have so much material I don’t have the luxury to read everything deeply
Then I usually ask: provide an overview of themes, agreements and disagreements. As I skimmed papers I have a rough idea already, this gives me more detail.
Then I question for specifics, example prompts:
- Which of these papers discusse point A?
- Do they all agree on said point?
- Please provide the direct quote before you explain.
- The last prompt is technically not necessary as there is a direct link to the paragraphs from each of the papers the information came from so you can verify immediately for accuracy.

Now at this point I had a discussion with a colleague:

Would that be cheating if you do this for an assignment?

The answer to this is a whole other blog post, from the necessity of internalised information building to develop critical thinking, to in another world I would pay research assistants to do the same work.²
https://www.makeuseof.com/notebooklm-settings-configure-notebook/

Perplexity

This is a new to me app coming with my phone and has some serious features, such as choice between the advanced AI models or it automatically chooses the model most appropriate to your question.

You can set it to computer mode, which is basically an agent setting, and deep research, which I still have to figure out. But so far this app seems the most useful for research. I had critical debates, push backs, and ‘have you thought about’. I only began using it and mainly because Gemini clammed down on data use and has become unreliable as research assistant.

Self-hosted platforms

I have been exploring to host a localised—data doesn’t leave my computer—AI, but right now I don’t have the headspace for this particular learning curve.

Let me know if you are using this, which platform, and how you got on.

I cannot stress how much it sucks for the ADHD brain to not have physical exercise to manage energy levels! Currently, see-sawing between extreme hyper focus and complete as in I fall asleep crashes. ↩︎
If I forget to write that other blog post prompt me. It’s ideas I am working I right now ↩︎