ChatGPT o1 preview + mini Wrote My PhD Code in 1 Hour*—What Took Me ~1 Year

Introduction

I recently had an astonishing experience where I tested a new version of ChatGPT, and the results left me both excited and a bit perplexed. I had been working on a coding project related to my PhD for nearly a year when I decided to put ChatGPT to the test using only the methods section of my research paper as a prompt. To my disbelief, it generated a functioning piece of code in a fraction of the time that I had spent painstakingly developing my own version.

Upon first glance, the generated code was remarkably similar to what I had created, despite the fact that I hadn’t shared any of my actual codebase with ChatGPT. The model managed to encapsulate the essence of my methods and replicate the expected functionality. My own code employed various functions—many of which I had crafted over many iterations—and here was ChatGPT doing it in about 215 lines.

I take pride in my methods section, as many papers have subpar methodology, and I thought mine stood out. Thus, it was gratifying to see that the details I put into my work paid off when ChatGPT used the information effectively. I initially thought about giving ChatGPT specific examples of my code but decided against it—opting instead to provide a high-level overview of my project.

Throughout this endeavor, I realized that while the code generated by ChatGPT appeared promising, there are still specific nuances and caveats that require human intervention. For instance, certain input parameters had to be established manually, as they could not be synthesized automatically. Moreover, any reliable results would necessitate further checks and adjustments—such as considering edge effects and the specifics of convolution operations.

Post -test, there were minor distinctions between ChatGPT’s output and my requirements. It did entertain analytic functions but could fall short in the areas that required intricate data manipulation by specialized software. Despite these shortcomings, I couldn’t ignore the potential it displayed in generating effective code from a textual description.

It leaves one to wonder whether the model accessed my GitHub repository to fetch my code. However, I took precautions to keep my code private. Nevertheless, if it did utilize my repository, the efficiency it showcased was impressive.

In conclusion, this experience has prompted me to reflect on the future of coding in research settings. The potential for AI to streamline such projects can be transformative, but it will also require scrutiny and delegation of actual coding and data modeling to ensure fidelity and accuracy.

Keyword

• ChatGPT
• PhD
• code generation
• methods section
• synthetic data
• input parameters
• convolution
• AI coding
• research

FAQ

Q1: How did ChatGPT perform the coding task so quickly?
A1: It used the description from the methods section of my research paper to generate functional code without needing to see my actual code.

Q2: Were there any limitations to the code that ChatGPT generated?
A2: Yes, some inputs had to be established manually, and there were additional steps for processing that weren't automated.

Q3: Did I share my actual code with ChatGPT?
A3: No, I only provided the methods section of my paper, so the model did not have access to my actual code or repository.

Q4: What challenges do AI-generated codes pose in research?
A4: While AI can speed up code generation, researchers must ensure that specific nuances and details are addressed to maintain accuracy and reliability.

Q5: Can AI really be a replacement for human coding in research?
A5: While AI can assist with coding tasks, it won't fully replace the nuanced understanding that humans bring to complex research projects.