I definitely agree on writing reports and code helping you think through the methodology and conclusions. It feels like 90% of the industry is trying to automate all their code writing and documentation at the moment.
Out of interst, how much have you adopted LLMs for these tasks? I worry if I am being too stubborn by just using it for debugging or writing simple templates / functions - it feels like for data science projects, writing "correct" code requires the context of the problem / user / use cases / conversations etc. which makes me hesitant to let an LLM make these decisions.
Hi Dom, honestly like most people I feel like I am still working it out. One thing I know for sure is that finding the line really matters. My current rule of thumb is “do I need to understand how this works?”. I am happy to vibe code simple utilities in Rust (and my Rust knowledge is pretty poor), and to farm out grunt work where all that matters is an output I can check. But if I need to understand something because I need to steer it, further develop it, or because it is part of some wider problem solving exercise then I am a lot more cautious - I sometimes use claude code for small steps but nothing more than that. Most of all I like the LLM to play the role of critic for the reasons I gave here https://glasseye.substack.com/i/185407626/the-dunghill. But tomorrow I might say something completely different!
I definitely agree on writing reports and code helping you think through the methodology and conclusions. It feels like 90% of the industry is trying to automate all their code writing and documentation at the moment.
Out of interst, how much have you adopted LLMs for these tasks? I worry if I am being too stubborn by just using it for debugging or writing simple templates / functions - it feels like for data science projects, writing "correct" code requires the context of the problem / user / use cases / conversations etc. which makes me hesitant to let an LLM make these decisions.
Hi Dom, honestly like most people I feel like I am still working it out. One thing I know for sure is that finding the line really matters. My current rule of thumb is “do I need to understand how this works?”. I am happy to vibe code simple utilities in Rust (and my Rust knowledge is pretty poor), and to farm out grunt work where all that matters is an output I can check. But if I need to understand something because I need to steer it, further develop it, or because it is part of some wider problem solving exercise then I am a lot more cautious - I sometimes use claude code for small steps but nothing more than that. Most of all I like the LLM to play the role of critic for the reasons I gave here https://glasseye.substack.com/i/185407626/the-dunghill. But tomorrow I might say something completely different!