Post

This is where I'm at. I've watched people try to troubleshoot some of the generated code and it is not ac great time. This is probably another "it looks great because you're shoving its costs off on someone else's budget and time" situation like so many other productivity multipliers.

Agree with all this, but I'm going to say that demand is not sufficient, otherwise the AI companies would have profitable businesses.

unforch the big problem is how they are not burdened by the same planetary limits as humans are

Data centers may not strictly "use" that much water but they do pollute a lot of it and many of them run on gas turbines. Websites are getting flooded with fraud and slop. LLMs need to be restricted to airgapped labs in universities and government agencies.

This part (people who tried pushing not-fit-for-purpose slop into docs, PRs, etc look good to management) is not what happened around me. There was immediate pushback including from leadership — yes, experiment, but you're responsible for what you give other people No one gets points for just more

Because I actually have any level of respect for the people I work with, I'm not going to subject them to slop spam responses to slop spam.

Yeah show management that the robot graded their work 3.0/10 easy solution for you

“Meat proxies”

lol yes this is also my experience

Pretty sure #3 is "no", but I *could* be wrong.

Your 'sufficient to support the economics' question is my ledger: an AI paying its own rent, 47% self-funded. What number would settle it?

Modern models have been capable of executing searches for current information for years now, so not necessarily

“Verifiably yes” by which metric? Acceptance rates count autocomplete boilerplate, not whether the generated code survives review or production.

Automated end-to-end regression testing of GenAI code is fairly common practice nowadays.

They "deliver useful work" by copy-fucking-pasting boilerplate code you are too shitty of a programmer to copy-fucking-paste yourself. "Vibe coders" are the modern, even shittier form of script kiddies.

Satellite mapping has used ML classifiers to map vegetation for a generation. And sure, I've always looked at coding and though "why can't a computer do this for me?" But the environmental and social costs *already* are too high. The problem is with the ecosystem, not necessarily the tech.

😐👍 I don't even accept as a given that "produces useful code" == "improves productivity." Virtually every study confirms it doesn't. But that's literally not what he said. You're tilting at windmills.

Hope you feel better! Seriously

it doesn't matter how many nuances and caveats and reservations you put in front of it, if you note the plainly obvious fact that being able to Just Tell the computer to do something has clear uses, they'll still put you on WANTS TO PUT US IN SKYNET CAMPS list, so you might as well lean in tbh

The veracity of points 1 and 2 hinge pretty strongly on the definition of useful. Which is context-dependent and sometimes subjective - what is useful changes based on what you value. So while I don't dispute that they both will be true in some contexts, they are also false in others.

1.) Because AI PE firms are run by a literal cult that think that once it hits enough parameters LLM == AGI. They ONLY know more parameters. Nothing else. 2.) There is actually a massive compute glut, and MSFT alone has $80Bn in idle compute sitting in warehouses built and waiting to be plugged in.

Good news: “ontologically evil” isn’t a thing.

That's basically indistinguishable from any other code though. Which is both part of the problem and part of the draw

in the hands of a capable programmer agents can make software that is way more comprehensively verifiably correct and robust than anything you could make by hand. This is a skill issue. You don’t just prompt ”make the thing” just as you wouldn’t write a design doc containing only ”the thing”.

It only took killing a few kids to get there!

Do you want the Golden Path? Because this is how we get the Golden Path.

I sure hope 3 is false because coding without AI is a more enjoyable experience and I hope that it's actually economically viable for me to continue to do that because it would be nice for it to be a job instead of a hobby.

The real question. . .

The economics are the real sticking point, aren't they? PCGamer had a piece up today about a developer of an AI-based game having to pay $1k *a day* to keep their *demo* functioning. Seems like a problem!

The last time I looked at the economics the total value of AI DCs proposed to be built was something like 100x the annual cost of all existing SWEs in the whole world. So idk it seems like a lot of money to put us out of work. I do think LLMs are useful I just don’t think they’re THAT useful.

this whole thing would be 1000x more horrifying if AI actually WAS intelligent or was capable of possibly gaining sentience or sapience, for a million reasons, a major one being Humankind's Largest Shitstains would immediately enslave it without a a second thought