From 1813e4c74acfe1b5d8a535e284ee91411ac2d895 Mon Sep 17 00:00:00 2001 From: Joseph Montanaro Date: Mon, 24 Aug 2026 06:23:41 -0400 Subject: [PATCH] work on llm post a bit --- posts/slam-dunk-uses-for-llms.mdx | 16 +++++++++++----- 1 file changed, 11 insertions(+), 5 deletions(-) diff --git a/posts/slam-dunk-uses-for-llms.mdx b/posts/slam-dunk-uses-for-llms.mdx index c16a5f0..d3b2563 100644 --- a/posts/slam-dunk-uses-for-llms.mdx +++ b/posts/slam-dunk-uses-for-llms.mdx @@ -28,16 +28,22 @@ But when I throw these situations at an LLM, it will usually find the syntax err And thank goodness for that, because tracking down a non-obvious syntax error is one of my least favorite ways to spend time. -### But they can still fall flat +### Even the failures are useful -Not that, again, LLMs are a magic bullet for debugging. In my experience, most bugs fall into one of three categories: +Of course, it's not as though LLMs are a magic bullet for debugging. I've had plenty of experiences where the LLM confidently asserted that it had "found the issue," and sometimes even that it had "confirmed" its findings, when in fact the problem was completely different. -1. Obvious bugs, where you know what's wrong as soon as the bug manifests, -2. Subtle-at-first-but-ultimately-obvious bugs, where the root cause is hard to track down but obvious once you see it, and -3. _Actually_ subtle bugs, the kind that you write up on your corporate blog after solving to reap those sweet sweet upvotes. +But even when this happens, it's not a total loss. A failed approach when debugging is still one possibility you don't have to investigate yourself, by hand. Not to mention that triumphantly explaining to the LLM how _wrong_ it is can sometimes cause me to realize what the problem _actually_ is, like a bank shot in pool. Or even (this has happened to me more than once) the very process of composing an explanation of the problem reveals the solution to me before I've even hit Send. + +Also, when it comes to debugging, Fable is just better. I know that size isn't everything, as the other AI labs are quick to point out, but it seems that it still counts for something. Several times I've thrown a problem at a lesser model, watched it go around in circles for a while, then switched to Fable and gotten the solution on the first try. ## UI Mockups +I'm absolute trash at coming up with a UI from scratch. It's funny because I usually have a pretty good idea about the abstract, high-level premises of a design, like what information should be presented and what actions should be available. But for whatever reason, when it comes time to sit down and bang out some HTML or whatever, I'm just _useless_. I can easily sit there in front of a blank page for an hour or more, trying one thing, getting annoyed and deleting it all, trying something else, before (frequently) getting annoyed and going off to work on something else for a while. + +So it's possible that my perception of an LLM's utility in this area is somewhat exaggerated. But I don't think that's the only factor. Consider: what's the biggest problem that most people have encountered with using LLMs? Their tendency to make stuff up out of whole cloth. What is _exactly_ what you want when you say "build me a mockup of such-and-such"? You want something that's made up of whole cloth. So really, you're just asking the LLM to do what it does best. + +In fact, you might want several somethings made up out of whole cloth, which is something for which LLMs can also be extremely useful. I don't know about you, but I have a strong tendency to become myopic once I've started building out a particular solution to a problem. Other solutions are less apparent to me the more time I've spent focusing on the one at hand. So taking a step back and asking an LLM to reimagine it from the ground up can help to break me out of the rut. Or I can start out by trying several different models on the same design, and pick the best.Admittedly I haven't actually tried this yet, so I don't know for sure that it would work. It's possible that their training is broadly similar enough that they will all end up with roughly similar solutions. Still, it seems worth a shot. Plus there's that "temperature" parameter that you can play with, if things seem too boring. + ## Brainstorming You can't uncritically accept everything it comes up with, of course, but if you're stuck an LLM can really help to get you unstuck sometimes.