Wednesday, October 4, 2023

SearchResearch Challenge (10/4/23): How might we best use LLMs for online medical research?

LLMs are all the rage these days... 

Illustration by DALL-E of a medical scientist doing research

Some studies show that office workers who use LLMs (e.g., Google's Bard or OpenAI's ChatGPT) are more productive... on the tasks that they measured.  (At least that's MIT's department of economics conclusion.)  

And it's pretty clear that LLMs and associated generative AI tools (see the illustration above) can come in pretty handy when given decent prompts.  

Our SearchResearch Challenge for this week is an open-ended Challenge.  For these Challenges, I usually know what the "right" answer is and how to find it.  

But this week, this really is a call for the SRS community to help identify ways of using LLMs to do a very specific kind of research--a kind that we care a lot about--medical research.  Here's the Challenge for the week: 

1.  What good advice can we give a person who wants to use an LLM for high-stakes research into a medical question?  We know people are doing this, and all of our ranting about LLMs won't stop them--but can we find ways of using LLMs and specific LLM/AI tools to improve the search process?  What would you say to people who are doing this?  (Aside from "be very, very careful.")  

Maybe the point of using an LLM in "medical research" mode is just to help us to find really great search terms to use in regular web-search.  I'm not sure. But if so, let's figure that out. Are there other ways of using LLMs to get better results faster?  Or will the search for truth in LLMs be a bust?  

I'm focusing on medical research because it is high stakes, it's not okay to just skate on past little errors and unfortunate phrasing.  

A big problem here is that most of us are not medical experts, so it's going to be difficult to validate what the LLM tells you.  

My suggestion: Try doing a few medical search tasks in a medical area that you know something about.  Nearly everyone has been forced to become a mini-expert in some medical area.  I'm a mini-expert in a couple of really unusual types of cancer (family members, not me!) when I had to learn enough to make good recommendations to them.  So I can explore what an LLM tells me about that with some confidence that I'll have a basic understanding of an esoteric area.   

I'll be leaving comments in the thread this week (more than usual) as I work on this question as well.  

Let's figure out what advice we can give to LLM users.  

When you write your comments in the thread, be sure to tell us which LLMs you've been using.  (There are a lot of them out there, all with very different behaviors.  Let's be sure to keep our results limited appropriately.)  

And be sure to include the search path you followed (which now will include LLM prompts).  Mention the dead ends you found and strategies that just don't seem to work.  Negative results are really important here!  

Keep searching!  


  1. Dan Brickley, who is a developer advocate for Google, writes (over on my LinkedIn thread) that:

    "Having api/plugin or whatever access to trusted repositories and sites seems an important piece of the puzzle."

    He also pointed to a really interesting paper on this topic: might be of interest.

  2. I was thinking about this all day. I'm still thinking.

    For example, what high-stakes means? I searched for the meaning of course, but, I mean the Challenge is to find for more than a flu, right? Like COVID maybe?

    How can we add those API, etc, when trying LLM?

    Another interesting point is: Already people is convinced (wrongly) that "Google MD" gives the answer, treatment, and more of all diseases. It's not Google fault of course. My question is: Now that LLM are supposedly more intelligent, this will be worst, right. I think your advice, Dr. Russell of being very, very careful is and will be more important every day.

    About LLMs, I only use Bard. I sometimes try Bing but just to try. Honestly, I don't use LLMs (creating correct prompts is not my forte. Even when tried creating images)

    1. I just read about this PDF

      From Eric Topol :

      "Encouraging results for large language model performance compared with surgeons for informed consent #AI "

  3. Refering back to our "-core" post Finding amazing words with LLMs, perhaps one strategy would be to ask an LLM for related terms for a medical condition. At very least, this might expand your working vocabulary and suggest alternative search strategies.

  4. One article that I found that seems relevant to both the professional and the citizen researcher

    Evaluating large language models on medical evidence summarization
    Liyan Tang, Zhaoyi Sun, Betina Idnay, Jordan G. Nestor, Ali Soroush, Pierre A. Elias, Ziyang Xu, Ying Ding, Greg Durrett, Justin F. Rousseau, Chunhua Weng & Yifan Peng
    npj Digital Medicine volume 6, Article number: 158 (2023) Cite this article

  5. looking backwards with bing - WOD or IOD

  6. some bits of AI working the genetic angle…

  7. old Blue: (this medical LLM search is producing a headache…
    & heart palpitations…) ;^[
    "Stress, exercise, medication or, rarely, a medical condition can trigger them. Although heart palpitations can be worrisome, they're usually harmless. Rarely, heart palpitations can be a symptom of a more serious heart condition, such as an irregular heartbeat (arrhythmia), that might require treatment."
    the business side:
    a LLM list: