Jeff AtwoodI have no idea what the source of this image is, but yeah, I Laughed Out Loud.
Bill the Lizard@codinghorror We should've just been nicer, I guess.
@lizardbill the funny thing is, the LLMs actually solve a HUGE problem for SO -- the endless repetition of the same questions over and over using slightly different words. I had NO idea how we would fix that, but now .. I kinda do?
aburka 🫣@codinghorror @lizardbill lol. Ask your mods if LLMs have been a net benefit.
Worik@aburka @codinghorror @lizardbill
> Ask your mods if LLMs have been a net benefit.
Ask me. Yes.
It is a tool that must be wielded by a hand. I do not think LLMs are going to be able to make reliable agents.
Unreliable agents are useful for some things, but reliable agents would be much more useful. Order of magnitude.
I want L5 automatic driving, I will not get it with current technology.
....liars
@worik @aburka @lizardbill I don't really like LLMs for code at all because such extreme precision is required in the language. That's not true of many other domains.
Jesse Skinner@codinghorror You're right.. I've come to realise they're really good for working with code if you minimize how much code they generate.. I paste in specs and have Claude Code look through the codebase and plan things for me, and I do almost all the implementation myself. Saves me a lot of digging and tracing and remembering, limits the hallucination, sycophancy and slop.
@JesseSkinner centaur model cc @AccordionGuy
@JesseSkinner you definitely SHOULD blog this, Jesse
@codinghorror thanks, I will!
@JesseSkinner lmk and I'll re-post here too
Jesse Skinner (@[email protected])
Do you love reviewing code other people wrote? Do you get a tickle of pure joy to find and criticize the mistakes and problems in sloppy code? Me neither. You know what I do love? I love pouring my creativity and insight and empathy into a project. I love designing architectures and solutions that actually make things better for users. I love getting in the flow state, cranking away at a problem, building brick upon brick until the creation comes to life. https://www.codingwithjesse.com/blog/coding-with-llms-can-still-be-fun/
@codinghorror oh, blog post link is here, curious what you think:
https://www.codingwithjesse.com/blog/coding-with-llms-can-still-be-fun/
BartyDeCanter@JesseSkinner @codinghorror Damn, I think you’ve convinced me to give LLM assisted coding a try again.
DNAvinci@BartyDeCanter
Agree.
Properly sandboxed, I can get Claude (opus, not regular) to work through a lot of the StackExchange questions that grad students never even typed in. Questions regarding the packaging and reproducibility elements of scientific software.
Also because the grad students published their expected output I can leave Claude alone with it to reproduce their setup.
This was previously burdensome and now catches me up to where I can do my actual work
@JesseSkinner @codinghorror
Agree.
Properly sandboxed, I can get Claude (opus, not regular) to work through a lot of the StackExchange questions that grad students never even typed in. Questions regarding the packaging and reproducibility elements of scientific software.
Also because the grad students published their expected output I can leave Claude alone with it to reproduce their setup.
This was previously burdensome and now catches me up to where I can do my actual work
@JesseSkinner @codinghorror
Adam Caudill@JesseSkinner @codinghorror I wonder if I’m just weird - but I do enjoy reviewing code. When I was a Development Manager, part of my evening routine was reviewing all the commits from the day and emailing out feedback and questions.
I wonder if that’s changed my perception of LLM coding tools, as reviewing code and providing written feedback is something I’m quite comfortable with. Plus, after leading dev teams, it’s become quite natural to offer detailed and precise instructions for changes to codebases I know well.
The perception and experience people have vary widely, and I’ve been fascinated with *why* for a while now. Some people see them as useless, others see real productivity improvements, and it’s often unclear what’s different.
wlach@adam_caudill @JesseSkinner @codinghorror Fwiw I enjoy reviewing code as well, but the big differences is that when I review another human's work, it's an opportunity for mentorship/up levelling as well as building shared understanding. I get none of that reviewing LLM output. AGENTS.md aside, every new session is a blank slate.
@wlach @JesseSkinner @codinghorror 100%. And there's a deeply frustrating since of futility from that - taking the time to provide detailed guidance feel useless, as you end up having to do the same thing again later.
Larger context windows have made that slightly easier to manage over time, but it's not scalable - the VRAM required for scaling up to the level needed to avoid that would be truly absurd.
@adam_caudill @wlach @codinghorror great points from you both, and that spurred me to rewrite the first two lines of the post, because I wasn't trying to dump on code review in general.
@JesseSkinner To be clear, I'm entirely willing to accept that I'm just weird! I do appreciate the edit though, I think that's a much clearer point.
Sean Reifschneider@wlach @adam_caudill @JesseSkinner @codinghorror "...every new session is a blank slate." I'm currently chewing on the idea: That's a good thing. You ever gone into code thinking "I wrote this, I know what's going on, I know how to make this change" but then later found you had missed some corner case? Like the other day I forgot to update the deb postinst. Coding your "mentoring" into the AGENTS.md but starting with clear context has been working very well for me lately.
@JesseSkinner After giving it a few days on a new project, I have to say that works pretty well. I added a few things that keep me honest and help me learn new languages.
- Do not offer to write code unless the user specifically requests it. You are a teacher and reviewer, not a developer
- Include checks for idiomatic use of language features when reviewing
- The user has a strong background in C, C++, and Python. Make analogies to those languages when reviewing code in other languages
@BartyDeCanter sounds great. I'm still fine tuning it as well.