Post

anthonymoser.github.io/writing/ai/h...

Except for the fact that Benjamin Franklin didn't actually say that, amen.

I think it still is a derogatory term. It’s just that since they’re thinking-outsourcers, outsourcing their thinking to the very same tool that’s meant to capture their attention and money, they may be a bit slow figuring it out.

Oh! THAT'S what vibe coding is! 😄

(E.g. me, who has been rejected because my methods weren’t compared to existing ones when there were no existing ones)

And a prediction machine of the most likely output, strictly based on previous publications, sounds like a very, very bad falsification tool.

5. ChatGPT has bullshitted them into believing that they're the next genius—but unlike previous generations of crackpot letter writers, they don't even pretend to understand their own "discoveries". They simply tell me that ChatGPT has a proof and they need my help getting the word out to the world.

I just made a sound in my throat at this

That's a weird and pretty bad example since there were a couple successful vibecoded mathematical proofs in the news recently... The key obviously being that formal(type-theoretical) mathematical proofs are easily verifiable whereas general science is not

I see that you have some of the same content as my spam folder. (Lately, I get people sending me "AI" copypasta pretending to have found aliens that have magic faster-than-light travel.)

It also involves (hopefully) very clear and precise definitions, the sort of thing genAI would be specifically bad at, as words in its training corpus would’ve defined differently than in a given research stream.

I don't know why anyone would do that. But then I know enough about coding for my own purposes though not enough to be hired for it. (I did Fortran on punch cards in college in the 70s. I've learned a smattering of other languages since as I've needed to.) I wouldn't trust any AI to write it.

If maintenance is even possible, which it often isn't

Don't get me started. No one seems to be able to reimburse me for my latest web infrastructure expenses because the person who gives permission to give permission to the person who gives me permission doesn't have permission to post to my account on the expense manager platform.

Some weirdo told me that AI, drones, and/or robots were going to take the fieldwork part of my job away and I was like, good luck with that when none of those things can do my desk work like financial reporting for grants.

"Trying to quickly make something that sorta works, where you can immediately sorta see if it sorta works and then sorta use it" is in fact a perfect description of how I & many other scientists write code. Very often enough to get a sense of whether you're headed in the right direction w a new idea

Its worse than that though. Science aims for outcomes that are verifiable and replicable. LLMs outcomes are by design inexplicable a non-replicable (identical inputs result in different outputs). All it ever produces is an illusion of meaning

Move fast and break things works for a tech startup with low risk and unimportant outcomes ("maybe my date will be a bad match") It doesn't work well for things with risk or important outcomes ("hey turns out vaccines actually worked, too late")

I might have the specifics of that wrong. I was just vibe-remembering, and my memory has probably been warped by the circumstances that led to him no longer being my boss.

a large fraction of what actually goes inside a physics paper can indeed be vibe coded

😂

I thought maybe you could run LLMs to look for causation and correlations so somebody could track it, you know, for things like astronomy calculations and other things. Not once did I think it should be used to do scientific theory.

Every decade we declare something "new thing anyone can do" and every time we rediscover that output quality tracks input understanding. This is "Excel will replace accountants" for a new generation. The floor rises, the ceiling stays computational.

Endless screaming into the void.

Riiight. I don't really know LaTeX nor do I care too much because ChatGPT works as a better Google for formulas and stuff (I am a very basic LaTeX user). So that's been super useful! But this is silly.

This is a chat box where you can try to yell at your paper until it is written.

The act of writing is fundamental to science, because the process is what makes people think. You cannot fully evaluate your idea until you are forced to write it down and defend it to yourself, and to the readers in your head. There is no shortcut for that. There is no way to make science easy.

Your being far to generous. There is no point to "vibe-coding". There's no point to anything GPT does. No understanding. No revelations. No connections. No provenance.