Zompist Bboard Again

malloc wrote: ↑Thu Jul 09, 2026 10:06 am they would have to embarrass themselves by making a movie they know is rubbish.

Bingo!

That simply doesn't make any sense from either a financial or branding standpoint, though. What would they have to gain by deliberately making a movie that they know will suck? So many arguments against the notion that AI is dangerous hinge on the assumption that the tech industry would mistakenly squander billions. Undoubtedly the industry has its share of fools, but fundamentally no industry can thrive for decades as computing has unless most people involved have some baseline competence.

malloc, while investors and corporate CEOs aren't stupid, they are not omniscient or perfectly rational, either. In fact they are often misled by their own greed, and many of them would rather burn a lot of money, saying to themselves that things are merely going through a trough and will turn out well in the end, than concede that they had bet on the wrong horse if their strategy fails.

WeepingElf wrote: ↑Thu Jul 09, 2026 1:10 pm malloc, while investors and corporate CEOs aren't stupid, they are not omniscient or perfectly rational, either. In fact they are often misled by their own greed, and many of them would rather burn a lot of money, saying to themselves that things are merely going through a trough and will turn out well in the end, than concede that they had bet on the wrong horse if their strategy fails.

Quite. For further proof of this, one need only consider just how many terrible movies the studios already turn out by their hundreds.

This is all a glaring example of what the late Alan Greenspan called "irrational exuberance", it should be remembered.

zompist wrote: ↑Wed Jul 08, 2026 3:02 pm psst don't tell malloc that the Care Bears are not actually played by bears

WHAT????? Sigh, another much-loved piece of my childhood turns to dust. What more reason is there to carry on?


... ponders emptiness of existence ...

Then how do you explain the success of generative AI in writing source code? Everyone agrees that qualifies as highly skilled labor, hence why software engineers have quarter-million dollar salaries. The criteria for good source code are also fairly objective, unlike with art and literature. You can easily discern whether an app functions as intended and avoids bugs. The fact that generative AI can produce functional software makes it hard to dismiss as merely a bullshit generator relying on our instinct to see patterns in noise as some have argued.

malloc wrote: ↑Thu Jul 09, 2026 10:06 am

zompist wrote: ↑Wed Jul 08, 2026 10:08 pm They can't do either.

Then what is this project using? Clearly they must have some working model for simulating actors or they couldn't attempt something like this in the first place. Otherwise they would have to embarrass themselves by making a movie they know is rubbish.

For an anticapitalist, you don't know much about capitalists. Most new businesses fail; entrepreneurs are optimists. And the promoter here is getting a shitload of publicity.

Here's a Wikipedia summary of reactions to their AI-generated short from last year:

Wikipedia wrote:Stuart Heritage of The Guardian described it as technically competent but "relentlessly unfunny to watch", with "sloppily written, woodenly delivered dialogue", and that Norwood's teeth kept "blurring into a single white block."[13] Joshua Wolens of PC Gamer wrote that Norwood's exaggerated mouth movements gave the impression "that her skeleton was about to leave her body",[14] while William Hughes of The A.V. Club wrote that the sketch's attempt at mimicking human body and mouth movements produced "such a hideous uncanny valley effect" that it gave them "a full-on case of the screaming fantods".[15]

As ever, the question is not whether you can produce video this way. The question is whether it's any good.

malloc wrote: ↑Thu Jul 09, 2026 3:21 pm Then how do you explain the success of generative AI in writing source code? Everyone agrees that qualifies as highly skilled labor, hence why software engineers have quarter-million dollar salaries. The criteria for good source code are also fairly objective, unlike with art and literature. You can easily discern whether an app functions as intended and avoids bugs. The fact that generative AI can produce functional software makes it hard to dismiss as merely a bullshit generator relying on our instinct to see patterns in noise as some have argued.

Generative AI isn't actually all that successful at writing code -- it often makes subtle mistakes that have to be teased out through careful code review and debugging -- much more so than code written by humans -- and furthermore the code it generates is often hard to understand and review in the first place.

malloc wrote: ↑Thu Jul 09, 2026 3:21 pm why software engineers have quarter-million dollar salaries.

Damn, if only.

The criteria for good source code are also fairly objective, unlike with art and literature. You can easily discern whether an app functions as intended and avoids bugs.

Tell me you've never worked as a developer without stating that you've never worked as a developer...

No, you can not "easily discern" if a program works. It's an enormous job, at least half of the work of a developer, to say nothing of the specialists (quality assurance engineers). Developers hate testing and code review almost as much as they hate writing documentation.

A person who uses AI to generate an entire program is not a person who does good testing, does good code review, or is good at finding bugs. It's someone who wants to take shortcuts— or is forced to do so by their boss.

There is in fact enormous pressure to use AI or be fired; so devs have to use it. Some report productivity gains; other report a loss as the AI's error's are hard to track down, and its code nearly impossible to read.

I'm out of the business now, thank the gods, but from the reports I've read, the AI's failure is in architecture. It can "write code", but it's like a junior dev who can do nothing but churning out code, who never thinks about extensibility or refactoring, whose code is unmaintainable.

A bad coder sees 100 tasks and writes 100 functions to handle them, never considering which could be handled by the same code, never thinking what happens when a new feature needs to be handled by all 100 tasks. A good coder is productively lazy: they group the 100 tasks into 10 or 5 or 2 tasks and just write that many functions. Guess which approach the AI takes.

zompist wrote: ↑Thu Jul 09, 2026 3:44 pm

malloc wrote: ↑Thu Jul 09, 2026 3:21 pm why software engineers have quarter-million dollar salaries.

Damn, if only.

That is certainly the figure I have always heard for salaries in the tech industry. I have consistently heard it touted as incredibly lucrative for everyone involved.

No, you can not "easily discern" if a program works. It's an enormous job, at least half of the work of a developer, to say nothing of the specialists (quality assurance engineers). Developers hate testing and code review almost as much as they hate writing documentation.

Suppose you are writing source code for an arithmetical calculator. The main criteria for functionality are whether the calculator gets the right answer for each operation and whether it respects the precedence of each operator. If your calculator says that two and two make five or that one times one is two, then clearly it wasn't coded properly. Now obviously most computer programs are considerably more complicated than a calculator with numerous features that all need checking. Nonetheless the same basic principle applies. You can easily tell if your browser loads webpages, if those webpages display correctly, and so forth.

There is in fact enormous pressure to use AI or be fired; so devs have to use it. Some report productivity gains; other report a loss as the AI's error's are hard to track down, and its code nearly impossible to read.

Sure but why this enormous pressure if AI doesn't work that well? The alternative to taking AI seriously is concluding that tech CEOs or even capitalists in general are complete idiots.

malloc wrote: ↑Thu Jul 09, 2026 4:51 pm

No, you can not "easily discern" if a program works. It's an enormous job, at least half of the work of a developer, to say nothing of the specialists (quality assurance engineers). Developers hate testing and code review almost as much as they hate writing documentation.

Suppose you are writing source code for an arithmetical calculator. The main criteria for functionality are whether the calculator gets the right answer for each operation and whether it respects the precedence of each operator. If your calculator says that two and two make five or that one times one is two, then clearly it wasn't coded properly. Now obviously most computer programs are considerably more complicated than a calculator with numerous features that all need checking. Nonetheless the same basic principle applies. You can easily tell if your browser loads webpages, if those webpages display correctly, and so forth.

Oh, you sweet summer child… even calculators are considerably more complex than you might think: see for instance https://chadnauseam.com/coding/random/calculator-app. (And that’s not even getting into the UI; I think I saw a whole other article on that alone.)

EDIT: Found it (and it’s well worth a read due to its clear and concise explanation of the sorts of problems programmers regularly face): https://lcamtuf.substack.com/p/ui-is-he ... alculators

Travis B. wrote: ↑Thu Jul 09, 2026 3:33 pm

malloc wrote: ↑Thu Jul 09, 2026 3:21 pm Then how do you explain the success of generative AI in writing source code? Everyone agrees that qualifies as highly skilled labor, hence why software engineers have quarter-million dollar salaries. The criteria for good source code are also fairly objective, unlike with art and literature. You can easily discern whether an app functions as intended and avoids bugs. The fact that generative AI can produce functional software makes it hard to dismiss as merely a bullshit generator relying on our instinct to see patterns in noise as some have argued.

Generative AI isn't actually all that successful at writing code -- it often makes subtle mistakes that have to be teased out through careful code review and debugging -- much more so than code written by humans -- and furthermore the code it generates is often hard to understand and review in the first place.

it kind of is though. one of the actually good things about llms is that they've made it much easier to build little applets that solve specific, concrete problems. a bunch of people in my company use a bunch of little app that do these small statistical operations: build a correspondence map, check for chisquare-significant differences in a certain type of table often used, scrape an app and report the findings. i'm not a programmer, but i'm a guy who knows how to program, but writing those little thingamajibs would have taken me weeks, and vibecoding them took me hours. plenty of people at the office have taken it to writing these small applets, and the work's been made easier for it.

of course i wouldn't vibecode production code for like an actual Application or whatever, but it's also not true to say the chatbots aren't good at making basically functional code quickly.

Torco wrote: ↑Thu Jul 09, 2026 5:08 pm

Travis B. wrote: ↑Thu Jul 09, 2026 3:33 pm

malloc wrote: ↑Thu Jul 09, 2026 3:21 pm Then how do you explain the success of generative AI in writing source code? Everyone agrees that qualifies as highly skilled labor, hence why software engineers have quarter-million dollar salaries. The criteria for good source code are also fairly objective, unlike with art and literature. You can easily discern whether an app functions as intended and avoids bugs. The fact that generative AI can produce functional software makes it hard to dismiss as merely a bullshit generator relying on our instinct to see patterns in noise as some have argued.

Generative AI isn't actually all that successful at writing code -- it often makes subtle mistakes that have to be teased out through careful code review and debugging -- much more so than code written by humans -- and furthermore the code it generates is often hard to understand and review in the first place.

it kind of is though. one of the actually good things about llms is that they've made it much easier to build little applets that solve specific, concrete problems. a bunch of people in my company use a bunch of little app that do these small statistical operations: build a correspondence map, check for chisquare-significant differences in a certain type of table often used, scrape an app and report the findings. i'm not a programmer, but i'm a guy who knows how to program, but writing those little thingamajibs would have taken me weeks, and vibecoding them took me hours. plenty of people at the office have taken it to writing these small applets, and the work's been made easier for it.

of course i wouldn't vibecode production code for like an actual Application or whatever, but it's also not true to say the chatbots aren't good at making basically functional code quickly.

Try developing for a real, non-greenfield project with tens of thousands of lines of preexisting code, actual coding standards, proper testing and code reviews, and so on, and tell me how far vibe-coding gets you.

Travis B. wrote: ↑Thu Jul 09, 2026 5:29 pm

Torco wrote: ↑Thu Jul 09, 2026 5:08 pm

Travis B. wrote: ↑Thu Jul 09, 2026 3:33 pm
Generative AI isn't actually all that successful at writing code -- it often makes subtle mistakes that have to be teased out through careful code review and debugging -- much more so than code written by humans -- and furthermore the code it generates is often hard to understand and review in the first place.

it kind of is though. one of the actually good things about llms is that they've made it much easier to build little applets that solve specific, concrete problems. a bunch of people in my company use a bunch of little app that do these small statistical operations: build a correspondence map, check for chisquare-significant differences in a certain type of table often used, scrape an app and report the findings. i'm not a programmer, but i'm a guy who knows how to program, but writing those little thingamajibs would have taken me weeks, and vibecoding them took me hours. plenty of people at the office have taken it to writing these small applets, and the work's been made easier for it.

of course i wouldn't vibecode production code for like an actual Application or whatever, but it's also not true to say the chatbots aren't good at making basically functional code quickly.

Try developing for a real, non-greenfield project with tens of thousands of lines of preexisting code, actual coding standards, proper testing and code reviews, and so on, and tell me how far vibe-coding gets you.

i know, i know, i'm not some tech billionaire trying to sell you a claude subscription. i explicitly said that i wouldn't use it for actual production code for an actual codebase. i get it's a tense issue but still

Torco wrote: ↑Thu Jul 09, 2026 7:09 pm

Travis B. wrote: ↑Thu Jul 09, 2026 5:29 pm

Torco wrote: ↑Thu Jul 09, 2026 5:08 pm
it kind of is though. one of the actually good things about llms is that they've made it much easier to build little applets that solve specific, concrete problems. a bunch of people in my company use a bunch of little app that do these small statistical operations: build a correspondence map, check for chisquare-significant differences in a certain type of table often used, scrape an app and report the findings. i'm not a programmer, but i'm a guy who knows how to program, but writing those little thingamajibs would have taken me weeks, and vibecoding them took me hours. plenty of people at the office have taken it to writing these small applets, and the work's been made easier for it.

of course i wouldn't vibecode production code for like an actual Application or whatever, but it's also not true to say the chatbots aren't good at making basically functional code quickly.

Try developing for a real, non-greenfield project with tens of thousands of lines of preexisting code, actual coding standards, proper testing and code reviews, and so on, and tell me how far vibe-coding gets you.

i know, i know, i'm not some tech billionaire trying to sell you a claude subscription. i explicitly said that i wouldn't use it for actual production code for an actual codebase. i get it's a tense issue but still

Some people actually like writing code, and conversely have little interest in bot-herding ─ and having an LLM write my code for me would take the fun out of it.

Some terrible non-AI news for a change: I spilled soup all over my chair and now it's ruined because the soup soaked into the upholstery. Honestly not sure what to do since I can't really afford a new chair, except maybe a cheap plastic one.

Torco wrote: ↑Thu Jul 09, 2026 5:08 pm one of the actually good things about llms is that they've made it much easier to build little applets that solve specific, concrete problems. a bunch of people in my company use a bunch of little app that do these small statistical operations: build a correspondence map, check for chisquare-significant differences in a certain type of table often used, scrape an app and report the findings. i'm not a programmer, but i'm a guy who knows how to program, but writing those little thingamajibs would have taken me weeks, and vibecoding them took me hours. plenty of people at the office have taken it to writing these small applets, and the work's been made easier for it.

I used to work for a company that did statistical programming, SPSS. And we had a team of statisticians who devised the formulas and checked the code to make sure it was correct. That's in addition to the designers and QA people, all separate from the developers. And believe me, there are a lot of gotchas in statistical programming. (And in what we consider simple arithmetic, as bradrn's link shows.)

I won't assume you didn't do testing... but what did it involve? Testing against a known piece of statistical programming? Running some calculations yourself? Eyeballing the code?

My two cents: a fair amount of programming involves problems that have already been solved. Most of the problem is already handled by the language or the framework; often it's a matter of using the right API correctly.

So in a good amount of case, it's a matter of locating the right piece of info in the API documentation or on Stack Overflow. It's not surprising the AIs are good at it.

I don't really know what kids these days are doing with Claude Code.
I use a LLM for small tedious tasks: lately because I wanted to check a certificate/private key pair. Of course the idea isn't to solve the problem myself but rather to use the openssl command correctly. The LLM gave me the proper command to use; which I could have located in the doc myself or dug up in a security guy's blog somewhere.
I also used it to work out a hairy SQL request (again, nothing that wasn't in the doc already) Ironically the initial idea for SQL was to have something close to natural language!

Not to answer for Torco, but I expect his use case isn't that different. The actual math is probably already worked out, in a python library somewhere, you just save time on tedious Googling/Stack Overflow crawling.

For the same reason, I expect the productivity gains touted by commercial LLM marketing are exaggerated. You save time on a (fairly tedious) part of the problem but the actual job is elsewhere -- in coming up with sane architectures, as zompist mentioned. It's not revolutionary.

I think the real revolution occured sometime in the interval between the time were programs were directly written in assembly, and circa 2015 when Spring Boot handles most everything for you.

Though in my view the real job is in figuring out what exactly the client is trying to do, what he actually wants and what he really needs. All three questions have very distinct answers. Also the client has no idea of what he wants, is unaware of what he needs so teasing out the answer is close to psychoanalysis.

Second thoughts on getting rid of coders.

I'm not saying there won't be a job loss (there is a productivity gain anyway).
But the programming part always made managers nervous.

They don't understand what happens; they're not developers themselvers. Or they're former developers, which is worse, because their experience probably involved BASIC on Amstrad CPCs, or punch cards. In my experience, coders aren't great at explaining what they do either.
Plus, they're usually optimistic by nature; they just don't understand that what in their mind would take a couple hours turns to take a whole week.

So it's not surprising that they expect that they'll finally be rid of the pesky undersocialized Mountain Dew addicts. It doesn't mean it will happen!