Many places are quite literally
forcing their software engineers to use LLMs. Complete with cursor/copilot is the ability to see usage statistics and surely at these companies these statistics will eventually be used as firing criteria.
I gave them a fair shake. However, I do not like them for many reasons. Code quality is one major reason. I have found that after around a month of being forced to use them I felt my skill atrophy at an accelerated rate. It became like a drug where instead of thinking through the solution and coming up with something parsimonious I would just go to the LLM and offload all my thinking. For simple things it worked okay but it’s very easy to get stuck in a loop. I don’t feel any more productive but at my company they’ve used it as justification to increase sprint load significantly.
There has been almost a religious quality associated to LLMs. This seems especially true among the worst quality developers and the non-technical morons at the top. There are significant security concerns that extend beyond simple bad code.
To me we have all the indicators of the maximum of the hype cycle. Go visit LinkedIn for confirmation. Unless the big AI companies begin to build nuclear power it will eventually become too expensive and unprofitable to run these models. They will continue to exist as turbo autocomplete but no further. The transformer model has fundamental limitations and much like neural networks in the 80s it’ll become more niche and die everywhere else. Like its cousins WYSIWIG and NoCode in 30 more years it’ll rise again like a phoenix to bring “unemployment” to developers once more. It will be interesting to see who among us was swimming without clothes when the water goes out.