[D] How to break free from LLM's chains as a PhD student?

A second-year PhD student's candid Reddit confession about becoming 'overreliant' on ChatGPT to write code has gone viral, exposing a growing crisis of AI dependency in academia. The student describes a year-long descent where the LLM progressed from automating 'boring parts' to generating core logic and templates, leaving them feeling they possess 'fake coding skills' and suffering from imposter syndrome. Compounding the issue is the perception that PhD advisors now mentally expect faster results, knowing their students use these tools, creating a pressure cycle that discourages foundational learning.

The post has sparked a massive discussion on strategies to reduce LLM dependency, with suggestions ranging from enforced 'no-LLM' coding sessions and using AI only as a debugger or rubber duck, to deliberately working on projects outside one's comfort zone without assistance. Commenters debated whether using LLMs is the new normal—akin to using Google or Stack Overflow—or a genuine threat to deep technical competency. The core tension lies in balancing the undeniable productivity boost of tools like GPT-4 and Claude with the need to retain the fundamental problem-solving and reasoning skills that define expertise, especially for those pursuing advanced degrees and research careers.