I was a researcher fifteen years ago. A PhD candidate doing Research for long days. I was swamped with many articles, annotations, emails, bookmarks, etc. When I found a citation manager tool, Mendeley, I felt so relaxed. It was like I had control over the process again. When I found a bookmark manager, XBookmark, I felt so productive (still have the bookmarks). They worked well for me at that time, and I finished my PhD program and got my degree. Episode 1 — Facing the Reality These days, I have an existentialist crisis. I see how AI research assistant tools have progressed. I was working with Scinito the other day, and I was shocked. I was trying to convince myself that this is only a tool to help in the literature review. Those who did a PhD know how tough it is to do a solid literature review. It is not a joke. You have to read over 100 articles, categorize them, understand them, and summarize them. If I say it took 3-6 months to do a solid literature review 15 years ago, I was not wrong. You heard me right, 3-6 months of your valuable lifetime. At first, I tried to convince myself that Scinito, or other similar tools, offer a marginal value to researchers. Sadly, or happily, I was wrong …  Not only can they do a literature review for you in a minute (I am sorry, my fellows, you heard it right), but they can peer review your articles. I would never forget how much I should have waited for my mentors and advisors to review my article, how many back-and-forth emails we had till things got up to an acceptable quality. Even after all these efforts, you get extensive feedback from peer reviewers in a journal to consider your article for publication. Or, you got rejected after 3 or 6 months only due to choosing the wrong journal to publish. These AI research assistant tools can enhance all these steps: reviewing your articles and selecting the most relevant journal for you.  Amazing. It is indeed amazing for researchers in this era, but I feel sad to see how much time I have spent on something that could have been done much easier and much faster. The interesting part is that this is not the end of the season. It is just the beginning.  This challenge is not just for researchers; it is also relevant to software developers. Tools like Cursor IDE have transformed how we build software dramatically. After my PhD, I started my career in engineering. So, I did lots of coding, testing, and so on. Today, I don’t need to read Stack Overflow to debug my code. I don’t need to spend time writing tests for my code. I no longer need to be an expert in React, HTML, or CSS to build a website. How much time did I spend on building sites in the past? I don’t want to think about it!  Episode 2 — Embracing the Reality Let me share the full half of the glass experience as well. This is super cool. I can ask AI research assistant tools to perform a semantic search within an extensive database. Something that we couldn’t have done before. It was just keyword matching. I could get updates about any topics or research questions in hours by reading the literature review that AI generates in seconds. I can write LaTeX code easily. I can reformat my paper to any guideline in minutes.  I am happy for researchers of our time. They can spend more time on creativity, problem solving, and, of course, their valuable life rather than doing unnecessary time-consuming tasks.  I am also happy for myself. I can write code in any programming language that I want. I can build websites without getting locked to Wix or WordPress. I can write any Python code that I need. I can optimize it and write a series of tests for it. WoW! It is super cool. The landscapes of coding, designing, researching, and everything are evolving fast. No matter how much people or organizations resist, Technology will find its way.  There is a catch here. The promise of building a website with one (and only one) prompt is not correct. I am telling this based on a very current experience. These days, I am working on a new website with a colleague, both of us are experts in software and AI. We didn’t even think about Wix or WordPress this time. We started using Curosr and its Agent experience with Claude-3.7-sonnet. The Cursor’s agent can generate the website structure in a second, but it fails when it comes to details. For example, when you want to align two different texts with each other, especially when one of them is static and one of them is dynamic, the AI can’t do it right. Basically, AI can do the website structure in a second, but can’t do the required details (details that you want to apply as a human on top of the prebuilt structure) as well as a UI design expert. That means, even though we don’t need to be experts in React or CSS, we must know the basics to intervene in the codebase when needed. Plus, we must know the concepts well enough to elaborate on them. If you can’t say it, AI can’t make it! I am not shocked by this weakness of AI models. They are built on the “wisdom of crowds” principle. It means they are built based on aggregating what’s most common, not emulating the intuition of an individual expert. This is rooted in their fundamentals. They are amazing in generality but suffer in specificity. In this short podcast, I explained a similar concept from a different angle: “The Erosion of Specificity.” Last Words I have been lucky to be part of the AI community. I am an AI architect with a solid plan to embrace this technology shift. But I am concerned about many other folks who can’t manage this change. This is not easy at all. If you have had an existential moment in your career due to AI, let me know. I may know something that helps. If I could share one tip here, it would be to “Learn the fundamentals, deeply.” You can (should) leave the repetitive, high-level, and generic tasks to AI, and spend your human creativity and expertise on the details to make your work/product shine. Follow me on YouTube if you want to hear more stories from the perspective of an AI architect: youtube.com/@AIUnorthodoxLessons/ The post An Existential Crisis of a Veteran Researcher in the Age of Generative AI appeared first on Towards Data Science.