Is editing ideas.json and prompt.json enough ?

[ibrahimptester]

hello,
I am not expert when it comes to coding and programming. I've understand till the installation and prepare scripts part after that what should I do if I want to write a paper on a different topic like bank and smes, I am thinking about deleting all the ideas in ideas.json and seed_idea.json and placing my own and modify the prompt.json but if I leave the experiment.py and plot.py as it is, will it generate a good paper ? as the code in nanoGPT is written for training ai models but I'm trying it on different topic.

[Krakaur]

I am not a developer, but based on what I have read, experiment.py and plot.py are designed for coding tasks (e.g., simulations) and generating visualizations (e.g., graphs). These scripts work well in fields like computer science or information technology. For disciplines like physics, mathematics, or chemistry, the code often requires modification to suit the specific needs of each field.

However, if your focus is finance, economics, or MBA topics, "experiments" and "plotting" take on a completely different meaning. The "core" of a paper in these fields revolves around:

1. Managing references (locating, downloading, reading, selecting, annotating, and processing).
2. Identifying and understanding the most relevant and useful ideas in those papers.
3. Creating a meaningful dialogue between the sources.
4. Ensuring proper style, formatting, and coherence.
5. Possibly downloading and processing quantitative data.

The advantage of working in these fields is that you don’t need an expensive GPU. However, the disadvantage is that you might need to do extensive work modifying the prompts. Additionally, citation and reference management are critical weak points that require considerable attention. For example, the Semantic Scholar API is currently facing challenges. The workarounds against paywalls are non ethically feasible, and web scrapping may fell under a grey area or be misused.

In my personal case, I’ve already built a collection of 100 downloaded papers, which I can "feed" into the model. However, I see that my roadblock is going to be the API cost.

We might overcome this issue with tools like DeepSeek V3 or by strategically working with less expensive, lower-capability models. These models could handle smaller, simpler tasks if structured appropriately. (e,g, it may be able to do 10x less, at a cost %1, so 10 times cheaper).