|
LLMs have taken much of the enjoyment out of coding for me, but I don’t think it has to be that way. I hope that as an industry we settle on LLMs being more like tools than human assistants. I think most of the engineers at my company are using LLMs as human assistants- most of them have agentic workflows set up and have premium subscriptions to Claude AND Gemini AND ChatGPT. Many have local LLMs running on their company MacBook Pros, but they can never manage to describe to me in plain English what those local LLMs are doing. I would compare local LLMs to Raspberry Pi clusters. They’re neat, and technically they can do stuff but they are incredibly impractical. I want so desperately to have the power of Claude Opus running locally, but we are incredibly, ridiculously, extremely far off (benchmarks are often misleading and nobody likes to talk about the paltry tokens/s they’re getting on their home rig). So the problem for me is that I’ve noticed a trend with my coworkers- they usually don’t have a formal CS background (e.g., Comp Sci degree), their resumes are unimpressive, and they have personal websites where anyone can download a copy of their resume. I haven’t seen this at any previous job. It’s just weird, man. Anyway, so my problem is when they are purported to be the best of the best at this company, and they’re chugging the LLM kool-aid, I’m struggling to figure out whether they’re really ahead of the curve and getting the drop on the rest of us, or they’re doing basically no work and waving through each other’s PRs with little to no oversight on what the LLMs are writing. The bush league nature of many of the bugs I’m seeing going all the way to prod makes me think it’s the latter. Meanwhile, these guys and their teams are completing an absurd number of story points each sprint. My team is dead last (and coincidentally we’re all pretty anti-AI for writing code). I’m waiting for the other shoe to drop. I’m waiting for the C suite to call our project a failure and lay everyone off. The product generally works, though I think if the devs were truly rockstars, things would be much smoother. It seems like LLMs are letting middle-of-the-road or mediocre devs appear to be rockstars purely on output speed. Nobody in this org cares if the output has 4x the bugs it should have, just that it was completed quickly. The main issue I think is that nobody- including the rockstar tech leads- knows how the system works. They don’t know the code. There’s a bug and it takes forever to track down. Don’t ask about unit tests. I think in a perfect world, these LLMs are treated like another tool rather than another person- meaning you and I still write a large portion of the code, and we read and understand every line- especially the lines generated by LLMs. It’s still a huge productivity multiplier for me to not have to remember how to do mundane things like write out arrays into CSV files- before I’d find a close implementation on StackOverflow and tweak to my needs. Now I can just describe what I want the CSV to look like and Claude Code generally gets it right on the first try. |