Currently, I have 3 main issues with AI, LLMs and so on. Generally speaking I am happy that I can use Claude Code and Codex to help building, fixing and maintaining software in my job. To be honest it is so helpful, I am less stressed bacause most of the time I can "find out" solution without asking collegues for help or I do not have spend a lot of time struggling myself. Yeah, if you understand what feeling it is that you know you can't solve problem then you will know what I am writing about.
But...
There are exceptions which make me bad feeling about AI, LLMs and so on:
1. Paid, subscribtion based payment model - you have to pay for tokens.
2. Local models are not "ready yet". You still have to own decent hardware and it will not be the same level of frontier models. I am big fan of local first, offline and "unlimited usage" compared to paid solutions. But probably I have to wait few or several years to get to that point.
3. Feeling "I did not do it myself". This is the most what could make bad feel. You "cheated". You did not write this great app or solve problem alone. You just did modern "copy and paste". You did not write or design this class, function alone by hand. And you know you probably could not do it without help of AI, LLM.
2. Local models are not "ready yet". You still have to own decent hardware and it will not be the same level of frontier models. I am big fan of local first, offline and "unlimited usage" compared to paid solutions. But probably I have to wait few or several years to get to that point.
This is why tech companies are working hard to make sure you can never build your own computer again.
The third point, on its own, doesn't bother me so much. What does bother me is the "clean it up for publication/PR/etc." phase. It was never especially fun in the first place, but cleaning up my own mess is something I can contend with, and you get a sense of satisfaction once you're done.
Cleaning up the slop from a faceless machine is just... not something I want to do. And even if you do get it done, the sense of satisfaction at the end is reduced.
I think the only way I can be truly "productive" with AI is to drop my bar for quality. For some cases that's easy (e.g. run-once-and-throw-away scripts), but for everything else I just can't do it.
Because a large portion of people are in this industry for the high salaries only, and always disliked the actual job of programming. Those people welcome LLM coding with open arms.
But...
There are exceptions which make me bad feeling about AI, LLMs and so on:
1. Paid, subscribtion based payment model - you have to pay for tokens.
2. Local models are not "ready yet". You still have to own decent hardware and it will not be the same level of frontier models. I am big fan of local first, offline and "unlimited usage" compared to paid solutions. But probably I have to wait few or several years to get to that point.
3. Feeling "I did not do it myself". This is the most what could make bad feel. You "cheated". You did not write this great app or solve problem alone. You just did modern "copy and paste". You did not write or design this class, function alone by hand. And you know you probably could not do it without help of AI, LLM.