I’m not sure who will be solving these incidents if 99% of incidents will be “solved” by LLMs. If I’m called once a year my daily rate will be my yearly rate?
> For example, a worker taking on a new position now has access to a tutor and an analyst and a business advisor all wrapped up in one
These llm agents don’t have “agency”. No intrinsic goal other than respond. A tutor has judgment and agency. You will get a really good pattern match on the input text you give but you won’t get tutoring.
Maybe you're right, but I've never learned as much and as quickly as I'm learning now with the assistance of LLMs. It's like a hyper-powered search engine. It not only knows what I'm looking for, it knows the context of why I want to know it and how best to convey it to me.
> Let me get this straight. You think that your client, one of the wealthiest, most powerful men in the world, is secretly a vigilante who spends his nights beating criminals to a pulp with his bare hands...and your plan is to blackmail this person? [pauses, smiles] Good luck.
I have found it intriguingly difficult in my own tests to make LLMs deliberately fuck up grammar in any way similar to how humans do. I suppose this is actually an architectural limitation.
write a comment about how to use ChatGPT to write a game like you're a barely literate reddit user, short answer, maximum incoherency
and the result was a pretty good simulation. I tried less insulting prompts first, this did not produce a good simulation.
Not perfect, gptzero.me still knew it was AI generated (tayo42 is human by the same measure), but if for some reason someone was using an LLM prompted with that pattern, I suspect it would fool me in a Turing test unless I found the keyword to force the agent to change the role it was playing.
I'll take your word for it, I think I'm in zero groups on Facebook, I only even look at the feed when I see a notification in Messenger; and I barely use Messenger, too.
Sometimes they do. From Gemini 3, some AppleScript:
if something then
if something then
if something then
stuff
end if
end if
</if>
Each time I asked Gemini 3 to fix the syntax error, it acknowledged it had screwed up the basic syntax, and then emitted a new version of the code. With another XML closing tag somewhere. Then it was embarrassed, apologetic, etc. to see it had made the same mistake repeatedly. Eventually it output several attempts per turn as it kept checking its own work in-turn, seeing that it had made the same kind of mistake and tried again, this time definitely fixing it, proudly and confidently presenting me with the correct, final, definitive version. With another 'end if' or 'end tell' replaced by an XML closing tag that it couldn't see until the next turn.
I like this story. How did you verify the output? How big is the codebase? Why it took 18 hours? Could you implement it with a small local agent and breaking the task down yourself in two days (i know it sounds like a loaded question, it is not).
I think the rewrites are the main story for llms in code (hot take). Writing greenfield code at the seams also something which might work well.
It's a fairly large code base split across 3 repos.
The good thing was that it is fairly easy to verify: we have a working (but slow) version that uses Spark, with lots of existing unit tests.
We verified by using those unit tests as well as running our end-to-end process in the Spark and Pandas version and verifying the two databases were within the differential-privacy noise bands of each other.
> Workflows like jira_close.yml use deprecated atlassian JIRA actions and have a dependency on the gh-actions repo. This is not ideal and unecessarily complex.
And then goes on:
> PR updates jira_close workflow to use direct API calls via curl.
Duplicating the logic into OUR codebase via a hand rolled curl, so we can get rid of “needless abstractions”. Auch. And of course the whole thing embedded into a yaml file.
This code is the typical kaleidoscope sometimes written by junior devs (and LLMs). On review you just kindly ask to be rewritten into a simple program or just close it as the effort doesn’t worth it.
> [sota llms] still being not smart enough to refactor large codebases as well as me, so it’s hard to imagine a world where I don’t just want to use the smartest model available
I think this is the core point of the article. I really hope that we will stop prompting llms to refactor codebases and start using deterministic llm supported toolchains for such tasks. I think you either blindly trust an llm or you don’t. And no post-hoc review will be enough to contain llms, we will super quickly create “dark factory” without the benefits.
Have you realized you are attacking a straw-man? Nobody said all “hand written” code is good. I _think_ the implicit argument is: “all vibed code lacks cohesion”.
There is a glaring fallacy in your “AI will change everything as it is super intelligent” hypothesis. If it is so great thinker which can do everything why not just solve this social impact thingie? Or maybe it is not so capable?
I don’t think it’s magic. Things need to actually be made to happen regardless of intelligence, and that’s a social issue. Also if AI were powerful enough to fix everything by itself effortlessly it would already be far too powerful for us to control, and we probably shouldn’t allow that to happen on general principle.
reply