The first time I asked something on Stack Overflow it took me like 20 minutes to figure out how to ask the question. I had to format the code somehow in that little input box, repro it, explain what I'd already tried, provide output etc. After all that work someone closed it as a duplicate. I was confused and unhappy to say the least but also, I didn't realize at the time, I was learning.
Needless to say, but nobody learns like this anymore. Now you go to ChatGPT or Claude and just get instant results (are they always right? who knows!). Although the era is gone now, I feel like there are some important lessons we can bring forward.
1) Writing the question WAS the lesson
Being able to state problems clearly is a HUGE part of our job in programming. LLM's are way more forgiving than other humans. Hell I don't even need to spell anything properly when I communicate with my LLM (I've never sent more typos in my life.) The 10-20 minutes just to craft the question was often a helpful exercise on what the hell was wrong. Even though my question sometimes got downvoted or closed I actually learned a lot in the process. I figured out what I was even confused about.
LLMs skip this part. You paste your error, vibe at the model until something works, and move on. You end up "solving" stuff you don't actually understand, which means you'll hit the same kind of bug next month and start from zero again.
2) Someone else has already solved this
SO's entire premise was community problem solving, and the underlying assumption that someone, somewhere has probably seen this issue before. But if through some rare circumstance it's not on SO, write your question clearly enough that the next confused person can find you. Often tedious but made everyone better at thinking and communicating complex problems. LLMs are the opposite. "Your problem is special (hell, you're special! you're also never wrong!) tell me about it and I'll happily fix it." Great for getting unstuck. Bad for learning anything.
3) The duplicate flag was a gift
Probably some of you are rolling your eyes, but "marked as duplicate" wasn't just annoyance and gatekeeping, it was a free pointer to the (sometimes) right answer plus a hint that you needed to get better at searching. The LLM will never tell you "this is a worse version of a better question, go read this." It just makes something up. Often worse than the dupe would have given you.
4) Juniors can't debug anything the model doesn't know
This is the part that scares me. I work with engineers fresh out of school who ship fast and can't debug for shit when the LLM is wrong. I often see people hacking away with an LLM for HOURS without ever looking at code. HOURS. They have no feel for what a good question even looks like because they've never had to write one. Getting unstuck without an oracle is a real skill and it's taught by never having had the oracle.
Conclusion
SO is a husk now, and we all know we're on a one way path. New questions have dried up. Nothing is being written down to replace what fed the LLMs in the first place. Eventually we're going to run out of seed corn and nobody's planting.
The models will keep getting better. The engineers using them won't. But sometimes I miss the duplicate flag, sometimes I miss the 2o minutes.
Comments