Post

Given the non-deterministic nature of LLMs, maybe sheep dogs are a better analogy than a table saw! bsky.app/profile/simo...

Simon Willison@simonwillison.net

That's entirely true: they are non-deterministic, which means you have to learn how to use them in a way that takes that into account Sheepdogs are non-deterministic too, they're still invaluable tools for someone who learns how to work with them

12 reposts118 likes

I've taken to talking about LLM as chef's knives -- very effective but you might take a finger off Tho sheep dog as a metaphor is pretty apt... border collies are super bright, super energetic, and occasionally have a spontaneous mental breakdown and aggressively go for someone

Love this observation: that table saw analogy really is an ink blot test for people's opinions of LLMs bsky.app/profile/yayp...

yay pineapple@yaypineapple.bsky.social

I went to your profile and read and re-read that table saw comment, and one can come away with several different interpretations! It's kind of like an ink blot test for one's thoughts about LLMs

Using an LLM is like having an adversarial agent data mining my research methodology and encumbering me with irrelevant distractions. It's an ADHD sheep dog that forgot to take its Ritalin and is being distracted by an orgy of squirrels and bouncing balls.

Agree. Are they perfect? No. Neither am I. Are they useful? For sure! Am I better with one in my tool kit? Absolutely. Are they overhyped? Hell yeah! Sadly, so is everything new in tech. I just don’t get the hate. Lean in. Learn how to use them. Get better. Lather rinse repeat. #kthxbye

I spent more than an hour strategizing and designing with Claude today. It was great and reminded me of having a couple of young programmers and a whiteboard. I imagined asking them how long to implement. Claude did it in ten minutes. I spent another half hour refining. Orders of magnitude for sure.

I just spent a few nights writing up my approach to trying to introduce some form of determinism to LLM-assisted coding workflows. Not a one size fits all of course. Would be interested in your thoughts. mariozechner.at/posts/2025-0...

mariozechner.atTaming agentic engineering - Prompts are code, .json/.md files are stateTreating LLMs as shitty general purpose computers we program with natural language. Because throwing shit at the wall wasn't working anymore.

Honestly, 'cheap homemade CNC machine' fits the bill quite well, but the number of people who might get that reference is vanishingly small. I suspect sheepdogs are much more reliable than cheap homemade CNC machines, which are, in turn, much more reliable than getting an LMM do do all your coding.