clarablog

i feel a sort of maniacal confidence about this, but also everyone around me irl who i think is really smart disagrees but also everyone of the people on twitter who i think are really smart kind of agrees. i am deciding to post it without editing the content apart from making it at least a bit legible after the crystal castles [5] and caffeine pill induced jitteryness wears off, as a sort of stream-of-thought rant.

there's a sort of frustration i get when talking to people irl about agentic ai. maybe i'm wrong and it really doesn't work for everyone and people just have vastly more complicated software problems than me and whatever...

but i can't shake the feeling that you're really all holding it wrong

i can't stop myself from starting conversations about this stuff and when i do, it usually goes like this

“but ai doesn't do that for me”

“are you using the paid version of claude or chatgpt as an agent on your computer”

“uhhh no”

“try it, it really is better than the chat interface”

“hm, but that's expensive/slop/not that capable”

and this sort of pleading goes on and i use codex and noone else seriously does. so what the hell are you all doing? one fellow student (who actually runs some sort of IT business so has more experience on this than me) who i had this discussion with said that he tried the agent and it didn't reduce the amount of work he had to put in. when i asked him what was taking up time when using it, he said that it was difficult to sandbox, because of the legal risk. we then discussed whether sandboxing the agent was a reasonable thing to do.

back then, the argument i landed on was “but any editor extension you install from some github repo carries the same risk for you as a developer, and if you're doing serious work where people would care to sue then you wouldn't have direct access to prod anyways”.

the statement i think was stronger that i was forming in my head back then but didn't finish say the conversation moved on was this: think of codex as a guy living inside your computer, ghost in the shell style. she's there to write your code. she's [1] kinda autistic but has read all of the code on the internet. ten to twenty years ago you would've paid a dude in india to do this. now you just have a guy inside the computer. isn't that great?

the guy inside the computer has weird preferences. he edits code by running unreadable python snippets, the same way some people edit code by running unreadable [2] vimscript snippets. it's kind of crazy to make him work inside of alpine:latest [3] because you worry he might delete your home directory if you don't let him ssh into your dev machine directly. he's working for you. [4] if you care about the project then don't “vibe-code” it but instruct the guy to make specific changes that you would make. don't insult him, if you see him going wrong then ask for follow up changes, look at things you care about in the output and in the code. just be normal about this.

i don't know what makes this not be the natural way to interact with ai for everyone. maybe my brain is just broken in a specific way that allow this sort of thinking to pass through unfiltered maybe there's just a certain kind of person who has less of the kind of cognitive facilities that distinguish themselves as an embodied person from an ai agent inside the computer.

maybe i'm going insane and the people around me actually have more difficult problems to solve and the technologically impressive projects that people online ship and all of the math the big labs are solving including the problems where mathematicians agreed beforehand that they are really really important don't actually matter for Real Work and don't imply that this stuff works and it's really just the case that everyone else is working on some really advanced enterprise:tm: software where you really need an adult biological human to keep it all in mind.

but also the people who live in cities that i think are actually cool and are not filled with despair and ideologically german institutions are the ones who are making money with this stuff, so maybe the ai people are just correct.

so my plea is maybe just to try adopting this model for a bit and seeing if ai still doesn't do the job if you just let it. use codex or claude code, and ask it to do the work and think of it as a guy doing it. i'm not asking you to give up all the “stochastic parrot statistical model that doesn't actually do any thinking” ed zitron stuff. i'm just asking you to try. or maybe i'm just going insane. i don't know. if it still doesn't do anything for you then i really don't know.

[1] i am uncertain about what gender assumption makes my case land the best here [2] more interactively created, and this is obviously not a good comparison because a long but minified python script is not the same as a vim command but also i think it is just ultimately fine that it uses these and people care to much [3] i didn't ask the coworker about the specific nature of his sandbox but i imagine it's something like “running codex in a docker container and copying the resulting code” [4] i believe auto review on codex or auto mode on claude code solve this, as in “makes it about as risky as using an unsandboxed command line shell yourself” (deleting your home is only one mistyped rm -rf ~/ away anyway) [5] the guy in this duo probably sexually assaulted his bandmate and by listening on spotify you are probably giving 0.003 $ per listen to a horrible person. i'm sorry, alice glass, i just can't get the music out of my head.

update 1: replaced the wired article with a cnn one that actually has the full text instead of only a synopsis update 2: placed the quote as such: “... any thinking” ed zitron stuff instead of “any thinking ed zitron” stuff