Programming thread

[ADHD Mate]

AI is much like regular development. I tell the computer to do something (compile, deploy, whatever), I wait for a bit, see whether it works, review, and then iterate.

I just put it on red, wait for a bit, see if I got some money back and then put on something else.

You aren't the best at abstract thinking? Dopamine system works on relative ease vs potential output. That's why people get addicted to things like scrolling. it's super easy to scroll feed, even if seeing something interesting is super rare.
Prompting is much easier than actually writing code with intent. So just asking LLM to do something, and if it fucks up asking again, to get something that works is addictive.
It's not even a question. There's plenty of people showing symptoms of addiction to AI coding.

The industry has already moved on. Most people have accepted that LLM usage will be part of their workflow.

Yet there are plenty anecdotal reports that code quality is tanking, domain knowledge is tanking, unhappiness is on the raise.
While I personally don't see that much gains anywhere.

The adoption of LLMs is still primarily driven by management. Anecdotally seniors/architects at my company dislike working with AI. You might say they are coping too. Which is fine. I am open to being wrong, but I don't think it has been proven yet.

There is also no point in trying to convince me with your, or anyone else, anecdotal evidence.
Especially with like I said before devs have been shown previously to overestimate gains. Whether the gains are real, and more important worth the price is not a settled matter.

 

[mnt_ms]

It depends on a game, but if we are talking about the greatest games out there - programming is the easiest part. You need good writers, designers, musicians. You are not going to write VtMB level characters with LLM, for example.

 

[SchizoDaemon]

You aren't the best at abstract thinking?

Don't be condensing, please.

Dopamine system works on relative ease vs potential output. That's why people get addicted to things like scrolling. it's super easy to scroll feed, even if seeing something interesting is super rare.
Prompting is much easier than actually writing code with intent. So just asking LLM to do something, and if it fucks up asking again, to get something that works is addictive.

It is a total false equivalence. I tried explaining that to you, and you just ignored it. It is incredibly frustrating when people do this. I've already explained my point of view on why it isn't addictive like gambling. I wasn't a gambling addict, but I get the thrill/rush you get from it.
I don't want to get into my own history, but I get really pissed off when people pretend that doom scrolling is like drug or gambling addiction.

That isn't what is happening when I or many others are using AI. It is simply a tool to get a job done.

It's not even a question. There's plenty of people showing symptoms of addiction to AI coding.

There are people that get addicted to Zyns, tea, exercising, sex, flossing their teeth, etc., etc., etc., etc. Many other people don't get addicted. Saying this doesn't mean anything.

Prompting is much easier than actually writing code with intent. So just asking LLM to do something, and if it fucks up asking again, to get something that works is addictive.

It isn't. I find it actually very boring, much like I find waiting for the compiler to build a large project boring or waiting for AWS to deploy the cloud formation change set. While the AI is editing one part of my code, I normally review other people's code or look at another ticket.

AI is just another tool I happen to use. The code I work on most of the time is very boring anyway.

I don't understand why you seem to want to constantly misrepresent what people are often doing.

Yet there are plenty anecdotal reports that code quality is tanking, domain knowledge is talking, unhappiness is on the raise. While I personally don't see that much gains anywhere.

Unhappiness happens because many people really shouldn't be in this industry. Management constantly doesn't value your contributions; they will throw away years of hard work on whim. There are many reasons why people are unhappy, and none of it has anything directly to do with AI.

There is loads of tangible progress made that is observable in many projects due to AI right now. I've read about 20 or 30 interesting projects where hobbyists and professionals have managed to solve things that would have maybe taken them years without AI, I think this month.

So you keep repeating that there aren't any gains. Yet I experience it myself. I see other people using it and being productive.

The adoption of LLMs is still primarily driven by management. Anecdotally seniors/architects at my company dislike working with AI. You might say they are coping too. Which is fine. I am open to being wrong, but I don't think it has been proven yet.

Is it? I think it is highly dependent on the company. Many of the developers I've spoken to have adopted it themselves. I have adopted it myself. At first I was using it simply as a better Stack Overflow to give me example code, e.g., "How do I do X in AWS CDK?" or using it to generate me mappings or other stuff.

I've gradually increased my usage as my knowledge of it has grown, and I understand what it is and isn't good at.

As for architects and seniors (I am normally in a lead/senior position), many of them are, let's say, "behind the times" or straight up retarded. This was true before AI became popular.

There is also no point in trying to convince me with your, or anyone else, anecdotal evidence.

The way you are conflating professional/expert opinion with an anecdote is a very annoying argumentation tactic. I will explain exactly why.

Let me give you some examples:

• I've repaired my own bicycles for years. I can tell you roughly how long it would take me to change an inner tube. I've done it loads of times (it is about 5-15 minutes depending on how tight the bead is on the tyre).
• I can tell you how long it would take me to change a clutch master cylinder on a LandRover Defender 300TDI (6-8 hours BTW), because I've done the job myself several times now.
• I can tell you the exact procedure to replace the front light cluster on a 1st-gen Opel Insignia and how long that will take (it is about 2-3 hours if you haven't done it before, 1 hour if you have).

Are these simply anecdotes? No. They are estimates based on my experience of doing the task as someone that has performed it several times.

Similarly. I've done what would have been epics in hours that would have taken days with an AI. I know roughly how long/how many story points it would take me to do a job without using the AI. Therefore, it isn't simply an anecdote. It is my professional opinion of someone that has been working in the industry for 20 years.

Especially with like I said before devs have been shown previously to overestimate gains. Whether the gains are real, and more important worth the price is not a settled matter.

I think some projects/orgs will see lots of gains. Others won't. I think it is highly dependent on the language, the infrastructure in the company and how management operates, e.g., I know AI has a lot easier time with, say, Go than C++. So you have to compare like-for-like in the same org and then find out where the bottlenecks are.

As for costs, you don't have to be running the latest version of Claude Fable and burning through hundreds of dollars of tokens to get reasonable results. I can get quite a lot done with some free or almost free models and Qwen 3.8, which I run locally.

I've said what I need to say, and we will leave it at that. Whenever I have these sorts of interactions on here about AI/LLM usage with you and others with similar opinions, you seem to be blinkers on, or I have utter nonsense thrown back at me like, "It is demonic" or whatever. I would rather people say, "I don't like using it," and be done with it rather than bring up all this bollox about addiction or whatever hyperbolic bullshit was trending on HN this week.

 

[moocow]

What kind of game would be hard to clone straight away? Even if you do something unusual, like performing pathfinding and collision detection on the GPU, a good LLM can still implement it for you.

Factorio. Go have your best/favorite LLM write me a new, clean-room implementation of a Factorio-like. Don't even worry about the Space Age DLC. Just the base game. Replicate the performance, scalability (of builds, that is), determinism (vital for multiplayer sync), multiplayer (the base game record is about 150 simultaneous users on one map in one session), and of course the gameplay, including all the machines, science research, pollution, enemies (biters), vehicles, trains, logistics and power. Remember maps are procedurally-generated (revealed via exploration during gameplay except for the spawn point at the center) and cap out at 1 million x 1 million blocks, at 50x50 tiles each. The target speed is 60 FPS and UPS (internal state updates per second) regardless of hardware. It should run well on toasters.

Let me know when it gets that done. The Wube fellas would probably love to talk to you if you can wrangle that out of an LLM.

 

[Whatevermancer]

Factorio. Go have your best/favorite LLM write me a new, clean-room implementation of a Factorio-like. Don't even worry about the Space Age DLC. Just the base game. Replicate the performance, scalability (of builds, that is), determinism (vital for multiplayer sync), multiplayer (the base game record is about 150 simultaneous users on one map in one session), and of course the gameplay, including all the machines, science research, pollution, enemies (biters), vehicles, trains, logistics and power. Remember maps are procedurally-generated (revealed via exploration during gameplay except for the spawn point at the center) and cap out at 1 million x 1 million blocks, at 50x50 tiles each. The target speed is 60 FPS and UPS (internal state updates per second) regardless of hardware. It should run well on toasters.

Let me know when it gets that done. The Wube fellas would probably love to talk to you if you can wrangle that out of an LLM.

That's yet another indie game with a long development cycle.

Each of which took over a decade to develop, which goes contrary to the popular game dev advice of making small games. Cause in the same time frame you can manage to release 10 different games and there's a chance one of them will become a hit and cover the development costs of the ones that flopped, instead of betting all on a single game that can turn out to be a failure.
Cause game dev is still a business.

 

[AirMail]

You are not going to write VtMB level characters with LLM, for example.

People will never get sick of making these "A.I will never get that much better" predictions, regardless of how many times they turn out to be wrong.

 

[mnt_ms]

People will never get sick of making these "A.I will never get that much better" predictions, regardless of how many times they turn out to be wrong.

Well, it's not a human, isn't it? It has no personality, no emotion, no point of view, and no grip on reality either. It's Globlobgabgalab the machine.

A good web search tool, though, all thanks to google being so shitty.

 

[Slurred]

There is this general trend in dev spaces of developer crying that they are now overworked because of performance expectations caused by LLMs. Which is interesting, because if they creating that massive gains I don't think it should be the case.

With eating, the hunger grows. Once you get used to AI coding, your ambition increases massively. Tools or features that would have been infeasible in the past can now be made in a fraction of the time. Anecdotally, I've made a lot of apps for myself that replaced various SaaS subscriptions and open source tools I was never quite happy with. The effort-to-reward ratio for most of these things made them not worth even starting before sufficiently capable coding agents, but now I've canceled a bunch of monthly subs and I have all these great personal tools that work exactly like I want them to.

Previous studies on AI productivity gains showed that developers think they are more productive, but in reality they were less productive. Obviously those studies were done on older models, and measuring productivity is tricky anyway so we might dismiss it.
But these things while still anecdotal still point that LLMs might not be that great.

It's precisely because they used older models. The study that made these claims was published in July 2025, and I fully believe it was true at the time. I tried vibe-coding a simple app in a framework I was unfamiliar with using Cursor around that same time and it was a total nightmare, just complete garbage. Also around that time, DHH went on Lex Fridman's podcast and told him that he still writes code by hand because the models suck at it.

But something changed at the end of 2025 with Claude Opus 4.5 and other frontier models released around the same time. Maybe it was because of improvements in the models themselves, or improvements in harnesses like Claude Code, or both working together, there's some disagreement about that. But in my experience and that of many others, that was the moment LLM coding jumped from being capable of one-shotting small scripts and doing semantic grep to being capable of building whole substantial projects that work as expected with minimal human code review. I re-attempted my previously disastrous vibe-coding experiments and the difference was night and day. It used to be shit but it got better.

All that while you accumulate cognitive and tech debt.

Anthropic itself released paper that showcases that using LLMs leads to impaired developer skill. A lot of productivity gains from AI require you to reduce reviewing code, and even if you do just briefly review you won't create internal model of your project as well as if you were solving problems yourself in the first place.
It's like tutorial hell but now with AI, it seems like you get it. But once something shit itself you are at disadvantage.

There's a stereotype of the non-technical idea guy vibe-coder who one-shots a barely thought out app and then immediately starts promoting it on social media and calling himself a startup founder, and then does the same thing twenty more times like he's pulling the lever on the one-armed bandit. But you don't have to work like that! Natural language is incredibly flexible and the models are capable of doing all kinds of things, not only endlessly inflating the line count.

You can (and should) specify architecture up-front, or on an unfamiliar project have your coding agent explain stuff with architecture diagrams and ask endless follow-up questions. You can get agents to delete dead code, refactor things, implement optimizations, test web interfaces in a browser session and take screenshots, or anything else you can think to express in a prompt. And you can do all of these things at whatever level of granularity you want to with whatever mix of manual implementation and code review suits you and your project. If anything, I feel like it's made me spend more of my time on the hard problems and decisions because it takes care of literally all of the drudgery.

This idea that any AI-developed system will inevitably hit a point where everything breaks and you have to fix it manually (or contract a heroic graybeard to do so) just doesn't ring true to me when AI has itself become a powerful tool for debugging. Modern coding agents are also really good at doing pedantic code review and finding bugs, including all the security exploits we're always hearing about. They can also be set up to monitor system health in production and alert you to issues and propose solutions, or you can use them as QA testers. Or if the project really is unsalvageable, you can just use the agent to rewrite it from scratch in a fraction of the time that would have taken before.

There's nothing inherent to the workflows or the tooling that mean you just have to endlessly slop out new apps and features with minimal thought. You can go into great detail with what you want and tinker endlessly. It's just a matter of working with these very flexible tools in an intelligent way.

People will never get sick of making these "A.I will never get that much better" predictions, regardless of how many times they turn out to be wrong.

One reason LLM tech has gotten good at stuff like programming and mathematics while still sucking at writing is that outputs for the former domains can be objectively verified. The test suite/Lean proof either runs correctly or throws an error the model can use to alter its approach. We don't really have a way to do that for "is this character well-written."

 

[UERISIMILITUDO]

This thing that I've staked my wealth on is inevitable, goyim!

Every big name in these corporations is Jewish. Something interesting I've read about in older books discussing artificial intelligence is one of the Jewish researchers, Edward Feigenbaum, amusingly comparing himself to the rabbi of Prague, the golem builder, in one quote. The rabbi was never meant to worship his golem as a god, however.

Anyway, there's a whole thread to discuss this neural network nonsense. Most programming conversations twenty years ago sucked, and now they're even worse because of this. Here's another OpenAI employee, Dean W. Ball, sharing his thoughts:
https://x.com/deanwball/status/2076029765631484042 (archive)

there comes a time in every ai policy professional’s life when they realize they have to read the talmud to make further progress, and for me that time was this week

John McCarthy and Richard Stallman are cool, though.

 

[canalCrab]

I did a stupid and started yet another attempt at making the game that's been in my head for an eternity. This time as a console application, where I am atm starting to build a framework for a TUI for it. Doing it in f# makes it surprisingly easy to implement a proper game tick loop

 

[Chiang Kai-shek]

>tfw you realize how terrible and over complicated your original program was

Better Sieve of Eratosthenes

(defun sieve(x)
  (let ((n (loop for y from 2 upto x collect y)))
    (dolist (i n n)
      (delete-if
                #'(lambda(z)
                    (and (>= z (* i i))(zerop (mod z i))))
                  n))))

A quicker solution I found, was to stop checking once the square of the prime numbers was greater than the number input. For instance, stopping at 7 when finding primes under 27 because the square of 7 (49) is greater than 27. Additionally, I finally get what memoization is and applied it. Even sped up the process by checking what the highest number in the cache was and, if the new number was less, fill up that lesser number’s prime list with number from the greatest until the prime numbers were greater than the lesser number. Closures are wonderful things!

Edit: Probably best I share the code itself

Spoiler: Prime Memoization

(defun easy-fill(lnum higherlst)
    "Generates the prime numbers of number from the list of a greater number"
    (do ((result nil)
        (locpri higherlst (cdr locpri)))
        ((> (car locpri) lnum) (reverse result))
        (push (car locpri) result)))

(defun prime-memo-2 ()
       (let ((highest 0)
         (cache (make-hash-table :test #'equal)))
         #'(lambda (args)
         (multiple-value-bind (val win) (gethash args cache)
           (if win
               val
               (cond ((< args highest)
                  (setf (gethash args cache) (funcall #'easy-fill args (gethash highest cache))))
                 (t
                  (setf (gethash args cache) (funcall #'primarch args)) ;;primarch is the sieve function pasted earlier
                  (setf highest args))))))))

Still has some kinks, but I think I'm heading in the right direction.

 

[moocow]

That's yet another indie game with a long development cycle.

So that's the cop-out? It baked for awhile with a small team?

What's the problem, LLM master? Get to work! Show them up! Reproduce it in a few days as you claimed was possible!

 

[Caesar]

People will never get sick of making these "A.I will never get that much better" predictions, regardless of how many times they turn out to be wrong.

LLMs will certainly reach a plateau. Whether or not there will be another technology to replace it or work alongside it is yet to be seen.

 

[Unintelligence Score]

LLMs will certainly reach a plateau. Whether or not there will be another technology to replace it or work alongside it is yet to be seen.

There is certainly some more efficient method than "consider the contents of all of the internet, including teenage blogs, boomer discussions and porn fanfiction, for the task of writing a shell script for listing files in the dist folder" waiting to be discovered..

 

[spacko]

ascii city also has a live feed

https://asciicity.live/

 

[Andy Soying Out]

Good video on how relying on AI for everything is decreasing your critical thinking skills required for programming:

Post in thread 'Vibe Coding & AI/SI Agentic Programming General'

Sep 30, 2026

Programming youtuber Kyle Cook just uploaded a video about how AI vibecoding is ROTTING your brain, and is ALOGGING faggot AI shill Theo.gg for encouraging vibecoding!
https://youtube.com/watch?v=ZumXpZzDsgo

• Andy Soying Out

 

[Least Concern]

Good video on how relying on AI for everything is decreasing your critical thinking skills required for programming:

Post in thread 'Vibe Coding & AI/SI Agentic Programming General'

Sep 30, 2026

Programming youtuber Kyle Cook just uploaded a video about how AI vibecoding is ROTTING your brain, and is ALOGGING faggot AI shill Theo.gg for encouraging vibecoding!
https://youtube.com/watch?v=ZumXpZzDsgo

• Andy Soying Out

https://youtube.com/watch?v=ZumXpZzDsgo

I can somewhat sympathize with struggling to make sense of LLM-written bug reports, though. In my own experience from having to make sense of vibecoded bug reports and PRs, there's two problems:

- The PR is written at the end of a chat conversation that contains context that I don't have access to, and especially which might be using terms for things which are not actually the ones in use by the rest of the project, so it's written using context I don't have access to and which might not even make sense to me in the first place, especially since I wasn't there at the start of it. I think this is especially a problem with non-technical people who might start the conversation with something like "fix the doohickey not turning on the thingamajig when I click on this thingy" rather than using the actual terms for things as used in the code.

- AI companies have a perverse incentive to keep LLM token usage high, so text that could have been "the Buy Now button is the wrong color for items in the Widget category" becomes some 700-word spiel starting with the fall of Byzantium to the armies of Suleiman or some shit.

Of course, when you are the vibecoder who has fallen into the habit of just wanting shit to be done regardless of how or why, that doesn't help things at all, so the guy's broader point still stands.

 

[Kosher Dill]

becomes some 700-word spiel starting with the fall of Byzantium to the armies of Suleiman or some shit.

I keep filing tickets on the fall of Constantinople to my government but they just mark it WONTFIX.

 

[DavidS877]

The PR is written at the end of a chat conversation that contains context that I don't have access to, and especially which might be using terms for things which are not actually the ones in use by the rest of the project, so it's written using context I don't have access to and which might not even make sense to me in the first place, especially since I wasn't there at the start of it.

I see something similar in a project to get vLLM running on the Intel B70 cards, sure there's a "here's how to reproduce it" with patches to a git hash, which isn't anywhere public, but the patches for that hash are also in the repo, and so on. All of it AI generated. Just tell your stupid AI I want a repo to clone and a dockerfile to build not something that has been iterated over 50 times by the LLM causing a dependency chain no human can follow. I asked free ChatGPT and after the third hop I told it "This is stupid, let's do something else."
Anyone wondering what that looks like: https://github.com/steveseguin/b70-...next-fp8-tp4-mtp1-exactgdn-b70-47tps-20260913

 

[King Of The Kiwis]

To you guys who work in the industry how can you guys code for 8 hours straight? I can barley code for more than an hour and a half and im a hobbyist. How do you guys increase your stamina?