I know this is only partly reasonable, but, this morning’s shower thought:

My work, as do most employers I know of in tech right now, requires us to use LLMs in some capacity. Otherwise, I would not involve LLMs in my work, since I find using them is generally an exercise in gold panning, even when I use a fleet of agents with a corrector / Socratic questioner. I find that use of AI slows my work down.

This is because LLM-generated code is, objectively, unoptimized slop and generally seems to take the naive partly wrong path even when it ‘works’. If I vibe code something and try to PR it without refactoring 80-90% of it, my CI tests and linters fail, the code mostly *doesn’t* do what I actually want, and seemingly does it the worst way possible within the parameters given. This is in response to doing my best to very specifically prompt for what I want. (If you are reading this and find yourself feeling the masculine urge to tell me to ‘just prompt better’, I will defenestrate you.)

This slop output is building on top of decades of people writing wasteful code because computing was ‘cheap’. The original lunar lander had the computing power of a fancy coffee maker. Now due to resource demand and lack of optimization and AI use, computing is expensive. If we wrote code as resource-consciously as we needed to when we had only fancy coffee maker level computing available, we wouldn’t need all these datacenters and a suggestion we might need ‘a buildout for the AI demand of the future’ would seem asinine