Post

deadsimpletech.com/blog/llms-ar... I agree that LLMs aren't inevitable in any strong sense but I find the article's claims that they produce mediocre results at best to be a bit dubious.

While it's in the nature of LLMs to be flattening, averaging machines, they've certainly accelerated the pace at which high quality work can be done. Recent maths work for instance -- not clean, but definitely high quality in the sense of boundary-pushing ("do a breakthrough")

mediocre work done at a high pace is way better than done slowly and in fact enables the high quality work to happen.

LLMs are like dishwashers. Sometimes you have to scrub stuff by hand, but I still wouldn't rent an apartment without one.

As an individual, LLMs are _ridiculously_ useful. Are they going to replace fleets of highly skilled software engineers? Can they do effective architecture and long-term maintenance? No, but they can certainly be used to offload the initial "make it work" phase

...and, with enough passes and a human-in-the-loop process, LLMs often produce high quality work. What counts as high-quality usually is polished, correct, maintainable code. It takes some fenagaling and multiple prompts, but LLMs can certainly produce that kind of output.

The article suggests that military might could be comparable to... number of software products released? But that seems like a tenuous comparison at best. Plus, Iris claims there hasn't been a major increase in products produced by major corporations... but what about Windows 11s many slop features?

And certainly for smaller orgs, if you want to compete on resources per proof of concept, an LLM will provide you something far cheaper far faster than human labor.