(2024-07-30) Udell Choosing When To Use - or Not Use - LLMs As A Developer

Jon Udell: Choosing When To Use (or Not Use) LLMs as a Developer. LLMs offer a level of augmentation that I wasn’t sure I’d live to see. They also bring an array of perils, many of which are ably documented in Baldur Bjarnason’s fiercely critical “The Intelligence Illusion,” including the one I’m addressing here: “We’re using the AI tools for cognitive assistance. This means that we are specifically using them to think less.”

So, when not to LLM? When You Don’t Want To Think.

I’ve written a lot of code over the years, but I am first and foremost a writer of prose. I’m also a spectacularly good copy editor.

LLMs make pretty good proofreaders, because they grok higher-order patterns. But I don’t rely on them for that purpose either. I want to keep exercising my proofing and copy-editing muscles

I’m wired differently when it comes to reading code. A bug caused by inconsistent spelling of a variable doesn’t jump out at me in the same way that misspellings in prose do.

Because reading code doesn’t come as easily to me as to others, I’ve tended to lean on LLMs for help. Now I’m reevaluating.

certain things, the LLM is a clear win. If I’m looking at an invalid blob of JSON that won’t even parse, there’s no reason to avoid augmentation

When You Don’t Want To Be Social

LLM-backed developer tools can serve as proxies for coworkers when you shouldn’t interrupt them. (rubber ducking)

When LLMs help us avoid unnecessarily interrupting others’ flow? Great. When they provide one more reason not to talk to people we should be talking to? Not great.

When You Need To Be Creative

Baldur nails the paradox: “The most productive way to use generative AI is to not use it as generative AI.” Instead, he argues we should use it “primarily for derivation, conversion, and modification.”

The hybrid solution I arrived at in my last column, Human Insight + LLM Grunt Work = Creative Publishing Solution, was, on the other hand, if not completely novel, then at least rare enough to not appear in the literature absorbed (at that time) by the language models.

It’s true that my use of LLMs to help implement the idea was, arguably, another form of creativity. For me, LLMs are software components, and my strongest superpower in the technical realm has always been finding novel ways to use and remix such components.

When the idea came to me, though, it did not appear on a screen in a conversation with Claude or ChatGPT: It appeared in my head while hiking up a mountain.