The question is not whether you can create software using LLMs - you can - most software is just boring CRUD shit.

[tante@tldr.nettime.org]

@map exactly. In a way accepting responsibility for the code one puts in front of people is accepting the connected care duties towards these people.

[mxey@hachyderm.io]

@map @tante even some pro LLM people see that: https://lucumr.pocoo.org/2026/2/13/the-final-bottleneck/

[ftranschel@norden.social]

@Nausipoule @tante I'd argue that instead (or, if you'd like: additionally), it is the terminal form of stochastic terrorism:

You will be randomly denied services, participation and dignity. Now isn't that quite a future.

[gklka@mastodon.social]

@tante @map That's just part of the truth. You can make wonderful, creative, unique software using AI. The thing is you have to specify what you want to achieve. If you don't give these goals to the AI then it will come up with some mediocre generic solution.

[elrohir@mastodon.gal]

@tante And I cannot even begin to emphasize how much *it will cost about the same despite it being of lower quality*.

Once credit entities realize that GPUs get obsolete very fast and five years down the line the early-mover needs to buy again just as much new processing hardware as a late-comer, they will stop subsidizing today's AI as a gamble to capture the market for tomorrow.

And then your 45-minutes-saving boilerplate machine will cost 5$ per run.

[kuzmandi@mastodon.social]

@pikesley @tante The Problem exists as long this "guys" are believing, that they can buy themselves a better world.

[nojhan@social.antigene.org]

@tante I mean… people accepted that for transports, agriculture, entertainments, even education and healthcare. Why stop here?

[nielsa@mas.to]

@tante Good take! But also, like "can you create software" is not really an accurate framing of what the hard part of software was.

Most people could "create software" by looking up a Hello world example. That wouldn't help them solve amy real problems tho.

LLMs produce software that *looks more like* it solves problems... but security, integrity, legality were kind of always implied parts of the problem.

Like, it takes a weird subtle reframing of the goal to make LLMs look at all useful.

[343max@mastodon.social]

@tante By now I’m pretty convinced llms can make it easier to produce high quality code then writing high quality code manually. Particularly because the AI is willing to do all the tedious, borring tasks that most developers are often to lazy for. Yes, it also makes it much easier to produce shittier code as well. (1/2)

[343max@mastodon.social]

Right now we are seeing way more of the latter because most people haven't learned yet how to produce good AI code and because the bad code sticks out while the good code blends in. But I'm convinced your underlying assumption “AI code = shitty" isn't correct. (2/2)
@tante

[cjk@chaos.social]

@343max @tante I think that's kind of the wrong question. Skill degradation and the moral implications (crawling of copyrighted material, climate, etc.) don't go away just because the generated code is good.

But I'm pretty sure you are aware

[vasilis@social.vasilis.nl]

@tante Then there’s the thing that we never had that software. Business has always accepted low quality products and services. So while I do agree with you, I’m afraid the people who run the software companies simply don’t care.

[343max@mastodon.social]

@cjk @tante Honestly I'm not sure about the skill degradation. I think there is a very high chance this is the same “new technology will make the youth stupid" panic that we have seen for centuries with every new technology. Also I really would like to see a deep analysis how much AI is hurting the environment. I don't trust Sam Altmans numbers but I also buy the “every prompt is burning down a small forrest" hyperbole.

[countholdem@mastodon.social]

@tante Yes, if I had Root Cause Analysis training - #BehavioralScience - in earlier education, I and more like me would've made more of a difference.

[alper@rls.social]

@tante No truly Germany has managed to give us great software over the past decades without LLMS.

[stefanfrede@mastodon.social]

@tante In my opinion, the problem is that most decision-makers don't understand the risks involved. When those in charge lack the necessary knowledge for software development, the empty promises of AI seem like an easy solution

[raymaccarthy@mastodon.ie]

@tante
Is that running or ruining?

[alessandro@mstdn.ca]

@tante

Entropy is somebody else's problem - they'll be comfortable on their yacht. We can't expect sociopaths to care about others, else they wouldn't be sociopaths.

[raymaccarthy@mastodon.ie]

@343max @tante
Then, Max, you have no understanding of LLM/Gen AI, or maybe of specifying requirements, designing system (modules, APIs etc) and then writing the code test & debug. If it's any size of project you need a team & management.
There is also documentation.
Actually writing the code is the easiest bit & the only bit the current LLM/Gen AI does, and does badly as it relies on code scraped from elsewhere & statistical shuffling of fragments.
Can't work. It's a technological dead end.

[hopeless@mas.to]

@tante Well, well done for admitting that demonstrably, the dog can play the piano. Now we are just talking about how well it plays.

FWIW these LLMs have no need for being consistent with what happened in a previous context. The same LLM, in a new context, will usefully critique and find and fix flaws in what it itself did in the previous context.

The "slop" aspect of LLM output seems to come from just going with what one context produced blind, when it can iterate as, eg, QA manager.