> Like it’s unmistakable to me how LLMs can basically up developer productivity to a much much higher degree than ever before. Yet we have plenty and plenty of people who can’t even take the middle ground and say it kind of helps. All kinds of developers everywhere saying LLMs are fucking completely useless. Which is mind bogglingly irrational.
There's not a single proper study showing this increase in productivity and just about every real developer I know finds very limited use in LLMs. They don't increase productivity "to a much higher degree". It's marginal, maybe 5-10% if you use them strategically in situations that are particularly suitable.
I decided to waste some time "for science" and implemented a feature twice, once by myself and once with Cursor. A feature that took me 4 hours to implement myself took 1-1.5 hours of planning + 1.5-2 hours of iterative agentic coding just to get it to meet basic functional criteria, and it would've taken me at least 2 more hours to review and refactor if I hadn't quit in frustration.
If I didn't care about long-term maintainability I could've finished it with AI in under 2 hours and I would've claimed a 100% productivity boost. I imagine that's what people do, prioritizing short term gains while taking on eye-watering amounts of technical debt, but trying to sell this as a productivity improvement is extremely naive.
LLMs are useful in very specific situations where the changes are trivial (small standalone snippets, straightforward changes in a larger codebase) OR long-term maintainability doesn't matter (one-off scripts). That's the middle ground.
There's not a single proper study showing this increase in productivity and just about every real developer I know finds very limited use in LLMs. They don't increase productivity "to a much higher degree". It's marginal, maybe 5-10% if you use them strategically in situations that are particularly suitable.
I decided to waste some time "for science" and implemented a feature twice, once by myself and once with Cursor. A feature that took me 4 hours to implement myself took 1-1.5 hours of planning + 1.5-2 hours of iterative agentic coding just to get it to meet basic functional criteria, and it would've taken me at least 2 more hours to review and refactor if I hadn't quit in frustration.
If I didn't care about long-term maintainability I could've finished it with AI in under 2 hours and I would've claimed a 100% productivity boost. I imagine that's what people do, prioritizing short term gains while taking on eye-watering amounts of technical debt, but trying to sell this as a productivity improvement is extremely naive.
LLMs are useful in very specific situations where the changes are trivial (small standalone snippets, straightforward changes in a larger codebase) OR long-term maintainability doesn't matter (one-off scripts). That's the middle ground.