What's your view on AI coding assistants?

[Arunbear]

I've not tried one yet

 27/40%

I'm not impressed

 20/30%

They are very useful

 18/27%

They are indispensable

 2/3%

67 total votes

[Fletch]

I'd almost want a middle option between 2 and three. I've moved from the "not impressed" camp back last year to "they're useful" (but not yet "very useful write all my code for me now clanker" though).

The free local ollama hosted (qwen 2.5 and 3, llama 3.3) that I've used so far are still lagging behind, but compared to 9-12 months ago there's noticeable improvements. "Commercial" offerings I've monkeyed with (Grok, GH copilot, Gemini Flash 2.5, Claude Sonnet) have similarly improved; I ran our interview coding problem through a couple of them and the ones I tried produced working not terrible code. Relatedly if I need to search for something depending on the context I'm going to use maybe 70/20/10 Grok/DDG/my local openweb with qwen.

For code assistants it feels like things have gone from complete slop for even trivial questions to mostly working short programs. At $work it's been decided to move to that ophidian competitor, and I find Grok pretty good for taking a detailed explanation "I'm trying to do X" and turning it into an ok/decent sample code or at the least providing me with enough search fodder to find what I really wanted.

I've also used GH Copilot (from emacs of course) to do some small utility scripts in that-other-language as learning practice and it was helpful producing slightly tweaked versions of a function which I'd already written.

So overall there's still a hype bubble for a lot of this, but if you're an experienced developer don't write things off. I think for experienced people especially moving into an unfamiliar language / domain they're at an interesting point where they can actually help you translate things into the new idiom / syntax / <handwaving />.

Edit: Not coding related, but nanobanana is pretty wild. Image stuff's come a long way from 13-fingered Will Smith consuming spaghetti through his 3 chins.

Edit 2: Coding related check out Aider which will use either an API key service (including OpenRouter) or local ollama hosted endpoints.

[adamsj]

++option 2.5

[choroba]

• They are somehow useful.
I'm currently writing a blogpost about my experience.

Update: Here it is: Using AI to Optimise the Calculation of Krippendorff's Alpha.

[talexb]

I think they do a pretty good job of analysis, and do an OK job on very narrowly defined tasks, but I wouldn't trust them to do any detailed work. Anything they generate is going to have to be examined very closely, and properly documented, before it can be used reliably. And if you have to spend that much time checking and fixing it, have you actually saved any time?

[LanX]

Supposing we talk about current LLMs
• They will produce a lot of slop.
• So called "Democratization" also means every baby can create dangerous code without understanding what he unleashed.
• What's unclear to me: Since maintenance is the biggest cost in a project, how will LLMs be able to cover that? Rewrite the whole code again and again?

[harangzsolt33]

So called "Democratization" also means every baby can create dangerous code without understanding what he unleashed.

Yes, this largely depends on how they use it. I think, everybody uses AI differently. When it comes to computer programming, I use AI to learn new things. For example, I ask questions like "How do I get the length of a string in Lua?" or "How do I open a file in binary mode in Python?" and Google AI gives an answer. It's the first thing that appears above the search results. Often I don't even click anywhere; I just read what was written by AI. I've found that sometimes AI quotes someone else word for word. For example, I click on the first result, and it takes me to a thread on Stack Overflow or somewhere where someone asks the same question and the answer that AI prints on the first page of the search result is the same that someone else wrote. So, AI is almost like a personal assistant that does research for me. I like it. As long as you ask very specific questions to which the answer can be found all over the web, there is no problem.

Sometimes AI tries to improvise. For example, I asked for a regex search and replace code in Lua, and it wrote something that didn't work. It was supposed to replace a substring within a string, and instead of replacing just that substring, the code written by AI replaced everything all the way up to the substring and the substring itself as well. So, that's not what I asked it to do, but whatever... I read up on it, and I understand now that Lua's version of regex is like a lame old man compared to Perl's. It's seriously disabled. Anyway, the point is that even AI couldn't figure out how to write a regex code for Lua that would work. So, when it comes to complicated tasks, you have to test it and double check everything to make sure it did exactly what you asked for, otherwise it's prone to make mistakes. But generally, I've found that AI makes life easier. Just don't ask complicated questions! Lol

[hippo]

there is no problem.

There is an absolutely colossal problem. Have you not noticed the unresponsiveness and at times complete unavailability of this site over the past 18 months or so? If you are using generative AI for anything at all then you are contributing to this problem.

The sooner the AI bubble bursts the better. I will not be paying for AI usage directly and I will not be using any AI system which is funding itself through advertising. We all have a responsibility to cut off the revenue from the perpetrators of the LLM training bot catastrophe if we want our web back.

[reisinge]

They're tools and as with all tools they are useful if you know how and when to use them.

[1nickt]

On a couple of occasions I have asked Google to produce some code (for example, JQuery JavaScript code to select/deselect all rows in a DataTables table using a singe checkbox input) and it has provided working code. But these are just snippets. So I can't speak to its ability to produce complex code.

I also find it to be of very limited use because at $work the vast majority of the "coding" we do consists of tiny incremental updates to our enormous codebase due to business rules changes. AI is familiar with neither the codebase nor the business rules behind it.

So for now it finds only an extremely limited role in my toolkit.

[roho]

As a tool they have their place, but you cannot take the human out of the loop.

[jdporter]

I haven't personally used a coding assistant yet, but I've asked CoPilot to write me some code on several occasions, and it has been an absolute boon. It has done in mere seconds what would have taken me hours the "old way". And sure, I have to check the code for correctness, but that's still orders of magnitude faster than writing the code from scratch. I (the human) become the verifier, in this mode.

ooh, here's a good article I found on the subject: Best practices for pair programming with AI assistants

[Discipulus]

• option 0: I avoid them
My very first interaction with AI was a question like: How does it costs asking to you in respect to a normal web search engine? AI: ..very difficult to answer because of bla bla bla.. me: try to spit out some number even if roughly inaccurate AI: my answer can be costier in energy by a factor around 20 and 200

I'd put a 150k cc engine under my kawasaki 750cc? No. My homeplanet is already burning, thanks.

That said I can imagine that an AI assistant can be useful for a professional programmer but at least with a grain of salt

Someone already mentioned software maintenence, but what about responsability? A bug in a game is comparable to a bug in nuclear plant?

Philosofically speaking I suspect AI lives in an ethernal present (nemesys of the Ethernal September ;) without being conscious of the context of time: my second (and last requested) interaction with AI was about a stupid oneliner of mines:

Me: what the following perl program does?  my $t = $^T; until ... { $t+=86400; ... etc

AI: the program computes days between January 1st ..bla bla..

Me: no! you are incorrect

AI: sorry I was wrong: ... days between January 1st..

Me: quitted as I dont want to train it

The AI did not abstract $^T and get fixed with January 1st (1970) while every human reads $t = $^T as now and now is relative because tomorrow will hold another value, so a human understand: days between the moment the program was executed and bla bla bla

L*

[CountZero]

I find them useful, but as with any other tool, much will depend on who is wielding the tool.

Recently, I have used AI to assist with a project I could not progress on for quite some time. Then I asked AI to write me some code. That code didn't work either, but gave me sufficient pointers to amend it into a working piece of code.

[adamsj]

My experience so far is that they're pretty good at doing things I can already do, but not so good at the things I don't yet understand. In particular, they aren't so hot for figuring out configuration errors.

[soonix]

Not exactly the same thing, but the same "big picture": One of my favourite AI quotes (as in "quotes about AI" as opposed to "quotes generated by AI") says

You know what the biggest problem with pushing all-things-AI is? Wrong direction.
I want AI to do my laundry and dishes so that I can do art and writing, not for AI to do my art and writing so that I can do my laundry and dishes.

(my source: https://indiepocalypse.social/@AuthorJMac/112178826967890119 by Joanna Maciejewska)

In her case, it's about what one likes to do vs. unwanted tasks, in (y)our case, it's about what we can assess because we have the expertise vs. where we'd need another expert who can assess things…

[bliako]

I feel it is even more wrong than that:

https://www.telegraph.co.uk/business/2025/10/12/why-western-executives-visit-china-coming-back-terrified/

“It’s the most humbling thing I’ve ever seen,” said Ford’s chief execu
+tive about his recent trip to China.
...

“I can take you to factories [in China] now, where you’ll basically be
+ alongside a big conveyor and the machines come out of the floor and 
+begin to assemble parts,” he says.

“And you’re walking alongside this conveyor, and after about 800, 900 
+metres, a truck drives out. There are no people – everything is robot
+ic.”
...

Other executives describe vast, “dark factories” where robots do so mu
+ch of the work alone that there is no need to even leave the lights o
+n for humans.

“We visited a dark factory producing some astronomical number of mobil
+e phones,” recalls Greg Jackson, the boss of British energy supplier 
+Octopus.

“The process was so heavily automated that there were no workers on th
+e manufacturing side, just a small number who were there to ensure th
+e plant was working.

...

“China has quite a notable demographic problem but its manufacturing i
+s, generally, quite labour-intensive,” he says.

“So in a pre-emptive fashion, they want to automate it as much as poss
+ible, not because they expect they’ll be able to get higher margins –
+ that is usually the idea in the West – but to compensate for this po
+pulation decline and to get a competitive advantage.”
...

worlds colliding. Literally! If you read the article you will see war mentioned there from that stupid, but honest, journalist.

no dishwashing in the trenches

[Random_Walk]

Its great for writing subroutines but try a larger progam and you will end up with a mess and a very long conversation with the AI trying to fix it.

I needed to take a flat list of items with and item number and the number of the parent item, and convert this into a tree structure in javascript. Not done much javascript and that was 10y ago so I asked AI(Gemini) and it gave me direct clean working code.

I've asked it to fix larger problems and spent too long debuging. It's also handy to generate tests and fake data. So I use it daily but build the framework of my programs first and get it to do the grunt work in well defined subs.

Cheers,
R.

[bliako]

Ignoring meta-physical theories trying to explain how humans are "unique", if they have a "soul", if they are different to other animals on this planet or indeed different to any aliens (the metric being some kind of turing-completeness test), I only see the difference between any AI and the human brain as quantitative, not qualitative. I.e. it is a matter of time and resources (if/when available) to reach the performance of the human brain. There is no other factor to stop this other than lack of resources (including: "there are not enough grains in the universe to fill a chessboard"/paraphrased).

I know the above is quite simplistic, but it looks to me that AI is on its way to parity.

What can stop this is Capitalism. Whatever Capitalism achieves in the short-term turns into ashes in the long-term. Of course what to the masses are ashes, can be diamonds for those "non-masses". But it's bubbling already given these massive, unrealistic contracts OpenAI does with Oracle for infrastructure, etc.

I can also answer this question by means of another question: what's your view on AI producing music, films, literature? What about flirting AI assistants? Or AI platonic partners? My answer: simply unacceptable.

It would be interesting to see if AI coding assistants subscribe to Perl's TIMTOWTDI or Python's TIO1WTDI/TINA. I would think the former given how eagerly the bots suck wisdom out of this Monastery.

Another thought: do you think Larry Wall's "the three great virtues of a programmer: laziness, impatience, and hubris" will be taken on-board by AI coding assistants? I think not, that would be political incorrectness for the current "west"-trained AI assistants or anti-conformist for the "east"-trained AI assistants - blasphemy (I can not compute!).

bw, bliako

[soonix]

AIs are not allowed to "I can not compute" - that's why they hallucinate 💭

[bliako]

nescio ergo (homo) sum!

that could be a good anti-bot captcha (yikes): ask them to describe a plum-bird. Or a monkey-tree. Or Another(?) Python hacker. lolllll

[starX]

Writing in my answer: it depends. I have had mixed results with them. In some cases, they have told me that certain libraries do things that I know, from my own personal experience, that they do not, and when I point this out, they politely apologized, offered me another solution that I knew would not work, and then politely apologized and referred me back to the original "solution" it provided. In some cases, I have asked them to integrate a solution into my existing code, and they have broken my code so badly that I've had to restore from my most recent commit to fix it.

But in other cases, they have been very helpful, and this includes both pointing me in the right direction for a tool or library I didn't know existed, combing through large quantities of data more quickly than I could write a regexp for, or even, in some cases, templating out a whole project in a language I didn't know very well to give me a good place to get started from. I suppose their utility is probably inversely proportionate to my own skill level in a given area or base competency in solving a particular problem. At worst, they can cost me an hour or two going down a garden path, but at best, they can save me literal weeks of work. So it depends.

[LanX]

> I suppose their utility is probably inversely proportionate to my own skill level in a given area or base competency in solving a particular problem.

Yes that's it basically.

When we solve a problem we intuitively try out strategies from solving other similar problems.

Without having much prior exposure to a special field LLMs can give you a start by finding or mixing known solutions they where trained on, but only if these solutions already exist.

They are basically sophisticated search engines with a better language interface.

Here is an anecdote where OpenAI embarrassed themselves.

It's highlighting that LLMs are not good at reasoning (thinking) but remembering (reproducing).

So as long as you ask things which require to reproduce learned knowledge they have a benefit. But asking them to solve new problems, not so much.

from https://techcrunch.com/2025/10/19/openais-embarrassing-math/ (emphasize added)

OpenAI’s ‘embarrassing’ math “Hoisted by their own GPTards.”

That’s how Meta’s chief AI scientist Yann LeCun described the blowback after OpenAI researchers did a victory lap over GPT-5’s supposed math breakthroughs.

Google DeepMind CEO Demis Hassabis added, “This is embarrassing.”

The Decoder reports that in a since-deleted tweet, OpenAI VP Kevin Weil declared that “GPT-5 found solutions to 10 (!) previously unsolved Erdős problems and made progress on 11 others.” (“Erdős problems” are famous conjectures posed by mathematician Paul Erdős.)

However, mathematician Thomas Bloom, who maintains the Erdos Problems website, said Weil’s post was “a dramatic misrepresentation” — while these problems were indeed listed as “open” on Bloom’s website, he said that only means “I personally am unaware of a paper which solves it.”

In other words, it’s not accurate to claim GPT-5 was able to solve previously unsolved problems. Instead, Bloom wrote, “GPT-5 found references, which solved these problems, that I personally was unaware of.”

Sebastien Bubeck, an OpenAI researcher who’d also been touting GPT-5’s accomplishments, then acknowledged that “only solutions in the literature were found,” but he suggested this remains a real accomplishment: “I know how hard it is to search the literature.”

[jdporter]

“Hoisted by their own GPTards.”

It makes me sad that they got the expression "Hoist with his own petard" wrong, as so many do. In Elizabethan English, "hoist" was the past participle of "hoise".

And to excuse it as being translation into Modern English doesn't fly, because the entire expression makes no sense in Modern English. It can only be understood by its original context, in Hamlet.

[bliako]

All in all, artificial and human intelligence teamed up to push the envelope of stupidity even further. Good news for this team is that Einstein conjured, I guess after ample experience, that the stupidity horizon is infinite. So plenty of ground there before this bubble bursts.

I also see how miniscule these "academic" types, for example the "mathematician Thomas Bloom", seem today. The maintainer of *THE* Erdos Problems website was unaware of 10 solutions and 11 more breakthroughs on the one and only field that his site specialises and is being advertised as such. Bloom gets away with just "I was unaware".

[syphilis]

I've had a few coding assistants over the years, but none of them were the product of Artificial Insemination ... as far as I know.

Cheers,
Rob