How does the ongoing generative AI boom affect my PhD (so far)? How I conduct research and go around my work duties has changed a lot in the last few years, especially in the previous few months. I want to share my thoughts, valuable tools, and insights in this post.
The private sector adapts innovation much faster than the public sector. That's how things work. Knowing that I was pleasantly surprised to see that Vrije Universiteit Amsterdam (VU Amsterdam) does innovate and, for example, has a 2-year MSc in artificial intelligence with several tracks, including "AI in Health". This speed of education adjustment to the changing reality of the world and emerging technology is impressive.
Nevertheless, I experience a limited amount of innovation from the top when it comes to my daily work. As a PhD, I
- perform a lot of manual labor
- on a project, nobody else in the world works on
- using an experimental setup assembled right here in the lab.
In other words, I am the worst target audience for a mass consumer-oriented product.
As much as I'd love to automate the experimental part of my PhD using machine learning combined with image processing, I don't have enough time in my Ph.D. for that. Imagine a setup that catches single DNA molecules, records all the force and fluorescence data from its interaction with various proteins, and performs the analysis. It sounds like a #PhDresearch project of its own or work for a whole company (LUMICKS tries to do something like that with the help of the scientific community).
However, I can change how I go around all the other parts of my PhD. Writing, coding, planning, meeting, and implementing new tools to make these things more straightforward and less time consuming.
I have learned about many AI-based tools through the AI club organized by my colleagues and supported by my supervisors; friends' recommendations; or just browsing the internet. Here are the ones I use the most in my daily work:
- Grammarly has been enhancing my writing for 2,5 years. It does a great job at making my text more convincing, engaging, and structured while tailoring the style to a particular audience, whether a LinkedIn post I write or an academic paper.
- Chat GPT 4 by Open AI became my indispensable programming buddy and saved me hours; otherwise, I had to spend digging through Stack Overflow or the various Python libraries' documentation. I've also used it for many other things: emails, table templates for my lab notes, abstract title inspirations, etc.
- Co-pilot is the most recent addition to my collection of tools. Thanks to the GitHub global campus, students and teachers can get it for free;)
- MidJourney comes in very handy when I need to generate images. If you've noticed, both scientists above are men. And so were all the scientists generated by MidJourney for other variations of these images (about 40 in total). Asking it to draw human portraits (below) doesn't produce the same gender bias, though there is a clear whiteness bias.
One of my primary duties and core skills as a PhD is to dig deep into the data and minimize the biases and errors in my analysis. Because of that, the conversation about the biases in the LLM and generative AI is fascinating to me. When the data fed to the model contains bias, the resulting model will be skewed accordingly. I have to account for that by modifying my prompts. For example, the blog post image took a lot of attempts to get right. Since Janni started the AI club, I wanted the central figure to be a woman. Prompting results were:
- "a group of scientists" produced exclusively male figures
- for "a group of scientists (men and women)" surprisingly, women still constituted less than <10%.
- "a group of scientists (women and men)" worked better, raising the women's fraction further. However, I had to select the image with the most women and do variations. This process got me to 33%
- "a group of scientists (women)" finally shifted the balance the other way. There were still some men, showing how embedded the bias is in the training data.
Just as with doing research, any user of this application must be well aware of the internal biases and potential causes of errors. As much as these AI tools promise a seamless and worry-free user experience, they require understanding their limitations and the dangers of overreliance on their outputs.
To wrap up this post, I'd like to share a service I have high hope for and many concerns at the same time: Elicit. This is an LLM tailored to scientific literature, which could give a depth response to your question, a literature overview, and valuable references.
See below the carousel of how it tackles the question, "Is red meat good for you?"
Yet again, so many papers published do not pass the quality of writing, research, and analysis standards. How would it ever filter them out?
