Today, I am wondering if Generative AI is one of the reasons that I started feeling aimless in my job as a researcher.

I work a lot with data in various forms and in the process of data analysis and preparation, I have started to rely quite a lot on the use of LLMs to do tasks for me that would otherwise be tedious and but represented a large share in what i was doing before. Last week, I was tasked to understand, structure, and organise a complex administrative data set with bad documentation, and unclear temporal coherency to make it available for analysis.

This would have been a challenging task some years ago. Maybe a week or two spent, focused on reading articles that have used the data set for other purposes, browsing documentation texts, looking up code bits on stackoverflow.

Instead, now, i bit by bit feed my instructions into an LLM trying to make sure it does the right things and provides the necessary documentation to keep the code sensible and human readable. But basically, I give a task, wait 20 minutes while the machine bobbles along, review what it had done, understand possible problems, set the next task, and then wait again.

In the in-between, I open a game of chess, or browse some articles (even interesting ones).

I can spend many hours like this, and usually finish the task quicker than i would have doing all the work by myself, or with an LLM of 3 years ago, where all code planning, and downloading was still done by hand, the LLM just helped me writing code blocks and syntax.

And I am wondering, if there is anything to do about this. Practically.

Because going back to the old world of doing 90 percent of the work (or even 100 percent), does not seem sensible from a pace perspective.

Yet, I also feel kind of empty while doing the task.

I am curious how other people have been in similar situation.

(I would like to propose to not make this a conversation about the ecological or ethical downsides of GenAI. They are real and relevant, yet, maybe this conversation can be more about the personal experience of LLM use)

  • Ohlalala@feddit.orgOP
    link
    fedilink
    arrow-up
    1
    ·
    1 day ago

    Thank you for the ideas!

    I am curious about your take on not guiding so much. Do you think this is possible?

    Because I do not really trust the machine to do a good job, and if I do not control what it does and clearly understand its process, I do not feel comfortable using its output in scientific publication.

    • funkydutch@feddit.nl
      link
      fedilink
      arrow-up
      1
      ·
      23 hours ago

      You can help it, good repo-level agents directives, personal directives, tooling checks, CI by another model, self-review. Bunch of stuff that comes packaged in those toolkits. But I think the power is in the personal customization not in those canned solutions. So you ask your best assistant what you asked me and how to ensure rigour for publication. Not that at I have experience with that. That’s tough. Sometimes its like “if you what you did is right then run this old data and prove it still works” kind of opportunities and we don’t notice.