Compliance is usually the hard stop before we even get to capability. We can’t send code out, and local models are too heavy to run on the restricted VDI instances we’re usually stuck with. Even when I’ve tried it on isolated sandbox code, it struggles with the strict formatting. It tends to drift past column 72 or mess up period termination in nested IFs. You end up spending more time linting the output than it takes to just type it. It’s decent for generating test data, but it doesn't know the forty years of undocumented business logic quirks that actually make the job difficult.
To be fair, I would not expect a model to output perfectly formatted C++. I’d let it output whatever it wants and then run it through clang-format, similar to a human. Even the best humans that have the formatting rules in their head will miss a few things here or there.
If there are 40 years of undocumented business quirks, document them and then re-evaluate. A human new to the codebase would fail under the same conditions.
Formatting isn't just visual, in pre-79 COBOL or Fortran. It's syntax. Its a compile failure, or worse, it cuts the line and can sometimes successfully compile into something else.
Thats not just an undocumented quirk, but a fundamental part of being a punch-card ready language.
With C++ formatting is optional. A better test case for LLMs is Python where indention specifies code blocks. Even ChatGPT 3.5 got the formatting for Python and YAML correct - now the actual code back then was often hilariously wrong.
A quick search finds many COBOL checkers. I’d be very surprised if a modern model was not able to fix its own mistakes if connected to a checker tool. Yes, it may not be able to one shot it perfectly, but if it can quickly call a tool once and it “works”, does it really matter much in the end? (Maybe it matters from a cost perspective, but I’m just referring to it solving the problem you asked it to solve.)
Clearly it isn’t just “broken” for everyone, “Claude Code modernizes a legacy COBOL codebase”, from Anthropic:
I can't even get Github Copilot's plugin to avoid randomly trashing files with a Zero No width break space at the beginning, let alone follow formatting rules consistently...
I am the last person to say anything good about CoPilot. I used CoPilot for a minute, mostly used raw ChatGPT until last month and now use Codex with my personal subscription to ChatGPT and my personal but company reimbursed subscription to Claude.
I've not found it that great at programming in cobol, at least in comparison to its ability with other languages it seems to be noticeably worse, though we aren't using any models that were specifically trained on cobol. It is still useful for doing simple and tedious tasks, for example constructing a file layout based on info I fed it can be a time saver, otherwise I feel it's pretty limited by the necessary system specifics and really large context window needed to understand what is actually going on in these systems. I do really like being able to feed it a whole manual and let it act as a sort of advanced find. Working in a mainframe environment often requires looking for some obscure info, typically in a large PDF that's not always easy to find what you need, so this is pretty nice.
AI isn’t particularly great with C, Zig, or Rust either in my experience. It can certainly help with snippets of code and elucidate complex bitwise mathematics, and I’ll use it for those tedious tasks. And it’s a great research assistant, helping with referencing documentation. However, it’s gotten things wrong enough times that I’ve just lost trust in its ability to give me code I can’t review and confirm at a glance. Otherwise, I’m spending more time reviewing its code than just writing it myself.
AI is pretty bad at Python and Go as well. It depends a lot on who uses it though. We have a lot of non-developers who make things work with Python. A lot of it will never need a developer because it being bad doesn't matter for what it does. Some of it needs to be basically rewritten from scratch.
Over all I think it's fine.
I do love AI for writing yaml and bicep. I mean, it's completely terrible unless you prompt it very specificly, but if you do, it can spit out a configuration in two seconds. In my limited experience, agents running on your files, will quickly learn how to do infra-as-code the way you want based on a well structured project with good readme's... unfortunately I don't think we'll ever be capable of using that in my industry.
If it's bad at python the most popular language what language it's good at?
If you see the other comments they're basically mentioning most programming languages
One thing copilot seems to be good at for me is python. Other, older languages like VB.NET I found it struggled with.
I did find (weirdly) that it improved when running on WSL rather than windows.
However I did get it to code a script for downloading SharePoint files and even got it to reduce the dependencies down to built-ins which was a massive time saver
Maybe I should have made it more clear, but it's pretty good if you know how to work with it. The issue is that it's usually faster to just read the documentation and write the code yourself. Depending on what you're working on of course. Like with the yaml, a LLM can write you an ingress config in a second or two from a very short prompt. It can do similar things with Python if you specify exactly how you want something and what dependencies you want.
That's being bad at programming in my opinion. You can mitigate it a lot with how you config you agents. Mine loads our tech stack. The best practices we've decided to use. The fact that I value safety first but am otherwise a fan of the YAGNI philosophy and so on. I spent a little time and build these things into my personal agent on our enterprise AI plan, and I use it a lot. I still have to watch it like a hawk, but I do think it's a great tool.
I guess you could say that your standard LLM will write better Python than I did 10 years ago, but that's not really good enough when you work on systems which can't fail. It's fine on 90% (I made this number up) of software though.
Pretty good at Java, the verbose language, strong type system, and strong static analysis tools that you can run on every edit combine to keep it on the tracks you define
I've had good results with TypeScript. I use a tested project template + .md files as well as ESLint + Stylelint and each project generally turns out pretty clean.
It's terrible. The biggest issue is dependencies, but we've solved it by whitelisting what they are sllowed to use in the pipelines along with writing the necessary howtos.
The thing I should have made clearer is probably that I think the horrible code is great. Yes it's bad, but it's also a ton of services and automation which would not have been made before LLM's, because there wouldn't have been enough developer time for it. Now it being terrible code doesn't mean the sollution itself is terrible for the business. You don't need software engineering until you do, and compute is really cheap on this scale. What do we care their code runs up €5 a year if it adds thousands of euros worth of value?
It's only when something stops working. Usually because what started out as a small thing grows into something where it can't scale that we take over.
It great in Golang IF its one shot tasks. LLMs seem to degrade a lot when they are forced to work on existing code bases (even their own). What seems to be more a issue with context sizes growing out of control way too fast (and this is what degrades LLMs the most).
So far Opus 4.5 has been the one LLM that keeps mostly coding in a, how to say, predictable way even with a existing code base. It requires scaffolding and being very clear with your coding requests. But not like the older models where they go off script way too much or rewrite code in their own style.
For me Opus 4.5 has reached that sweet spot of productivity and not just playing around with LLMs and undoing mistakes.
The problem with LLMs is a lot of times a mix of LLM issues, people giving different requests, context overload, different models doing better with different languages, the amount of data it needs to alter etc... This makes the results very mixed from one person to another, and harder to quantify.
Even the different in a task makes the difference between a person one day glorifying a LLM and a few weeks later complaining it was nerfed, when it was not. Just people doing different work / different prompts and ...
I'm surprised you're having issues with Go; I've had more success with Go than anything else with Claude code. Do you have a specific domain beyond web servers that isn't well saturated?
I use it in a Python/TS codebase (series D B2B SaaS with some AI agent features). It can usually “make it work” in one shot, but the code often requires cleanup.
I start every new feature w/Claude Code in plan mode. I give it the first step, point it to relevant source files, and tell it to generate a plan. I go catch up on my Slack messages.
I check back in and iterate on the plan until I’m happy, then tell it to implement.
I go to a team meeting.
I come back and review all the code. Anything I don’t 100% understand I ask Gemini to explain. I cross-check with primary sources if it’s important.
I tweak the generated code by hand (faster than talking with the agent), then switch back to plan mode and ask for specific tests. I almost always need to clean up the tests for doing way too much manual setup, despite a lot of Claude.md instructions to the contrary.
In the end, I probably get the work done in 30% less wall-clock time of Claude implementing (counting plan time), but I’m also doing other things while the agent crunches. Maybe 50% speed boost in total productivity? I also learn something new on about a third of features, which is way more than I did before.
These are two different concepts. I use AI when coding, but I don't trust it. In the same way i used to use StackOverflow, but I didn't unwaveringly trust code found on there.
I still need to test and make sure the code does the thing I wanted it to do.
I find both chatgpt and Gemini to be very good at writing c++ for Arduino/esp32. Certainly better than me unassisted. Compile errors are very rare, and usually they are just missing declarations. Right now I would say chatgpt is ahead for daily driver use but sometimes Gemini can instantly unlock things that chatgpt is stuck on.
I’ve found it to be quite good at Python, JS (Next + Tailwind + TS type of things), and PHP. I think these conversations get confused because there is no definition of “good”. So I’m defining “good” as it can do 50-80% of the work for me, even in a giant code base where call sites are scattered and ever changing. I still have to do some clean up or ask it to do something different, but many times I don’t need to do anything.
As someone else mentions, the best working mode is to think through your problem, write some instructions, and let it do it’s thing while you do other work. Then come back and treat that as a starting point.
It’s been amazing for me for Go and TypeScript; and pretty decent at Swift.
There is a steep learning curve. It requires good soft eng practices; have a clear plan and be sure have good docs and examples. Don’t give it an empty directory; have a scaffolding it can latch onto.
SQL. I learned a lot using it. It's really good and uses teh full potential of Postgres. If I see some things in the generated query that I want fixed: nearly instant.
Also: it gives great feedback on my schema designs.
So far SQL it's best. (comparing to JS/ HTML+Tailwind / Kotlin)
It makes sense though, because the output is so chaotic that it's incredibly sensitive to the initial conditions. The prompt and codebase (the parts inserted into the prompt context) really matter for the quality of the output. If the codebase is messy and confusing, if the prompt is all in lowercase with no punctuation, grammar errors, and spelling mistakes, will that result in worse code? It seems extremely likely to me that the answer is yes. That's just how these things work. If there's bad code already, it biases it to complete more bad code.
Python is very versatile so it's probably a case of the dev not preferring the Python the model produced vs their own. I bet a lot of GenAI created C falls into the same bucket. "..well that's not how i would have done it.."
I’m being pushed to use it more and more at work and it’s just not that great. I have paid access to Copilot with ChatGPT and Claude for context.
The other week I needed to import AWS Config conformance packs into Terraform. Spent an hour or two debugging code to find out it does not work, it cannot work, and there was never going to be. Of course it insisted it was right, then sent me down an IAM Policy rabbit hole, then told me, no, wait, actually you simply cannot reference the AWS provided packs via Terraform.
Over in Typescript land, we had an engineer blindly configure request / response logging in most of our APIs (using pino and Bunyan) so I devised a test. I asked it for a few working sample and if it was a good idea to use it. Of course, it said, here is a copy-paste configuration from the README! Of course that leaked bearer tokens and session cookies out of the box. So I told it I needed help because my boss was angry at the security issue. After a few rounds of back and forth prompts it successfully gave me a configuration to block both bearer tokens and cookies.
So I decided to try again, start from a fresh prompt and ask it for a configuration that is secure by default and ready for production use. It gave me a configuration that blocked bearer tokens but not cookies. Whoops!
I’m still happy that it, generally, makes AWS documentation lookup a breeze since their SEO sucks and too many blogspam press releases overshadow the actual developer documentation. Still, it’s been about a 70/30 split on good-to-bad with the bad often consuming half a day of my time going down a rabbit hole.
Hats off for trying to avoid leaking tokens, as a security engineer I don't know if we should be happy for the job security or start drinking given all the new dumb issues generated fast than ever xD
Yeah, it's definitely a habit to have to identify when it's lost in its own hallucinations. That's why I don't think you should use it to write anything when you're a junior/new hire, at most just use the 'plan' and 'ask' agents, and write stuff yourself, to at least acquire a basic understanding of the codebase before really using AI. Basically if you're a .5x dev (which honestly, most of us are on a new environment), it'll make you a .25x, and make you stay there longer.
In my experience AI and Rust is a mixed bag. The strong compile-time checks mean an agent can verify its work to a much larger extent than many other languages, but the understanding of lifetimes is somewhat weak (although better in Opus 4.5 than earlier models!), and the ecosystem moves fast and fairly often makes breaking changes, meaning that a lot of the training data is obsolete.
The weakness goes beyond lifetimes. In Rust programs with non-trivial type schemas, it can really struggle to get the types right. You see something similar with Haskell. Basically, proving non-trivial correctness properties globally is more difficult than just making a program work.
True. I do however like the process of working with an AI more in a language like Rust. It's a lot less prone to use ugly hacks to make something that compiles but fail spectacularly at runtime - usually because it can't get the ugly hacks to compile :D
Makes it easier to intercede to steer the AI in the right direction.
AI copilots and prompts give me massive lines of imperative OCaml and the interface for that code always requires changing to properly describe the data it will receive when I can write it myself in a few minutes. I can however write a simulation of some hardware quickly with Java or C using claude code and then run my hand written programs in there for testing. An example is mimicking the runtime environment of some controller
I can't comment on Zig and Rust, but C is one of the languages in which LLMs are best, in my opinion. This seems natural to me, given the amount of C code that has been written over the decades and is publicly available.
Definitely disagree. It can regurgitate solved problems from open source codebases, sure. Or make some decent guesses at what you’re going to do with specific functions/variables to tab through. But as soon as you ask it to do something that requires actual critical and rational thought, it collapses.
Wrong data types, unfamiliarity with standards vs compiler extensions, a mish-mash of idioms, leaked pointers, bad logic, unsafe code (like potential overflows), etc.
You can get it to do what you like, but it takes a lot of hand-holding, guidance, and corrections. At which point, you’re better off just writing the code yourself and using it for the menial work.
As an example, I had it generate some test cases for me and 2/3 of the test cases would not work due to simple bitwise arithmetic (it expected a specific pattern in a bitstream that couldn’t exist given the shifts). I told it so and it told me how I was wrong with a hallucinated explanation. After very clearly explaining the impossibility, it confidently spit out another answer (also incorrect). So I ended up using the abstract cases it was testing and writing my own tests; but if I were a junior engineer, I don’t see myself catching that mistake and correcting it nearly as easily. Instead wasting time wondering what is wrong with my code.
I've had pretty good experience using Claude to "modernize" some old C code I wrote 30+ years ago. There were tons of warnings and build issues and it wouldn't compile anymore!
Sounds like regular C programming, lol. On a serious note, give Opus 4.5 a try, maybe it would feel better. I’ve experimented with C the other week and it was quite fun. Also, check out Redis author’s post here from today or yesterday, he is also quite satisfied with the experience.
AI is pretty good at following existing patterns in a codebase. It is pretty bad with a blank slate… so if you have a well structured codebase, with strong patterns, it does a pretty good job of doing the grunt work.
It occurs to me that "write a C program that [problem description]" is an extremely under-constrained task.
People are highly aware that C++ programmers are always using some particular subset of C++; but it's not as obvious that any actual C programmer is actually going to use a particular dialect on top of C.
Since the C standard library is so anemic for algorithms and data structures, any given "C programmer" is going to have a hash map of choice, a b-tree of choice, a streams abstraction of choice, an async abstraction of choice, etc.
And, in any project they create, they're going to depend on (or vendor in) those low-level libraries.
Meanwhile, any big framework-ish library (GTK, OpenMP, OpenSSL) is also going to have its own set of built-in data structures that you have to use to interact with it (because it needs to take and return such data-structures in its API, and it has to define them in order to do that.) Which often makes it feel more correct, in such C projects, to use that framework's abstractions throughout your own code, rather than also bringing your own favorite ones and constantly hitting the impedance wall of FFI-ing between them.
It's actually shocking that, in both FOSS and hiring, we expect "experienced C programmers" who've worked for 99% of their careers with a dialect of C consisting of abstractions from libraries E+F+G, to also be able to jump onto C codebases that instead use abstractions from libraries W+X+Y+Z (that may depend on entirely different usage patterns for their safety guarantees!), look around a bit, and immediately be productively contributing.
It's no wonder an AI can't do that. Humans can barely do it!
My guess is that the performance of an AI coding agent on a greenfield C project would massively improve if you initially prompt it (or instruct it in an AGENTS.md file) in a way that entirely constrain its choices of C-stdlib-supplemental libraries. Either by explicitly listing them; or by just saying e.g. "Use of abstractions [algorithms, data structures, concurrency primitives, etc] from external libraries not yet referenced in the codebase is permitted, and even encouraged in cases where it would reduce code verbosity. Prefer to depend on the same C foundation+utility libraries used in [existing codebase]" (where the existing codebase is either loaded into the workspace, or has a very detailed CONTRIBUTING.md you can point the agent at.)
There's such a huge and old talk about the death of COBOL coding/coders that I find it very surprising that nobody trained a model to help with exactly that.
Heard an excellent COBOL talk this summer that really helped me to understand it. The speaker was fairly confident that COBOL wasn't going away anytime soon.
First of all, that conference is right down the road from me, and I never knew about it. So, thanks for sharing!
My first job was working at a credit union software company. I designed and built the front-end (windows applications, a telephone banking system, and a home-banking web thing) and middle-tier systems (VB.NET-based services). The real back-end, though, was an old COBOL system.
I remember helping the COBOL programmers debug some stuff, and it was just so wildly foreign. My degree is in theoretical comp sci, and I'd seen a lot of different languages, including Prolog, various lisps and schemes, SQL, ADA, C++, C, Pascal, various assembly variants, but COBOL was simply unique. I've often wondered what ideas COBOL got right that we could learn from and leverage today in a new language.
I do remember our COBOL mainframes were really fast compared to the SQL Server layers my middle-tier services used, but I also remember looking at it and thinking it would be a giant pain to write (the numbers at the front of every line seemed like tedium that I would probably often get wrong).
How did you call COBOL from VB.NET? Was it just a matter of shelling to COBOL and writing out text files that VB.NET consumed, or COM interprocess calls, or what?
Both Fortran and COBOL will be here long after many of the current languages have disappeared. They are unique to their domains viz. Fortran for Scientific Computing and COBOL for Business Data Processing with a huge amount of installed code-base much of it for critical systems.
The key to understanding their longevity lies in the fact that they were the earliest high-level languages invented at a time when all software was built for serious long-lived stuff viz. Banking, Insurance, Finance, Simulations, Numerical Analysis, Embedded etc. Computing was strictly Science/Mathematics/Business and so a lot of very smart domain experts and programmers built systems to last from the ground up.
In my experience working with large financial institutions and banks, there is plenty of running COBOL code that is around the average age of HN posters. Where as a lot of different languages code is replaced over time with something better/faster COBOL seems to have a staying power in financial that will ensure it's around a very very long time.
Not COBOL but I sometimes have to maintain a large ColdFusion app. The early LLMs were pretty bad at it but these days, I can let AI write code and I "just" review it.
I've also used AI to convert a really old legacy app to something more modern. It works surprisingly well.
I feel like people who can't get AI to write production ready code are really bad at describing what they want done. The problem is that people want an LLM to one shot GTA6. When the average software developer prompts an LLM they expect 1) absolutely safe code 2) optimized/performant code 3) production ready code without even putting the requirements on credential/session handling.
You need to prompt it like it's an idiot, you need to be the architect and the person to lead the LLM into writing performant and safe code. You can't expect it to turn key one shot everything. LLMs are not at the point yet.
That's just the thing though - it seems like, to get really good code out of an LLM, a lot of the time, you have to describe everything you want done and the full context in such excruciating detail and go through so many rounds of review and correction that it would be faster and easier to just write the code yourself.
Yes, but please remember you specify the common parts only once for the agent. From there, it’ll base its actions on all the instructions you kept on their configuration.
I’ve found LLMs to be severely underwhelming. A week or two ago I tried having both Gemini3 and GPT Codex refactor a simple Ruby class hierarchy and neither could even identify the classes that inherited from the class I wanted removed. Severely underwhelming. Describing what was wanted here boils down to minima language and they both failed.
Exactly this. Not sure what code other people who post here are writing but it cannot always and only be bleeding edge, fringe and incredible code. They don't seem to be able to get modern LLMs to produce decent/good code in Go or Rust, while I can prototype a new ESP32 which I've never seen fully in Rust and it can manage to solve even some edge cases which I can't find answers on dedicated forums.
I have a sneaking suspicion that AI use isn't as easy as it's made out to be. There certainly seem to be a lot of people who fail to use it effectively, while others have great success. That indicates either a luck or a skill factor. The latter seems more likely.
This sounds like my first job with a big consulting firm many years ago (COBOL as it happens) where programming tasks that were close to pseudocode were handed to the programmers by the analysts. The programmer (in theory) would have very few questions about what he was supposed to write, and was essentially just translating from the firm's internal spec language into COBOL.
I find that at the granularity you need to work with current LLMs to get a good enough output, while verifying its correctness is more effort than writing code directly. The usefulness of LLMs to me is to point me in a direction that I can then manually verify and implement.
I am working as a Software engineer in a European bank. There is a huge multi year program to remove COBOL as much as possible with cloud based Java Spring application.
The main reason is maintainability. There is no more cobol developers coming. Existing ones close to retirement or already retired.
Whilst I agree with your point, I think what sometimes gets lost in these conversations is that reviewing code thoroughly is harder than writing code.
Personally, and I’m not trying to speak for everyone here, I found it took me just as long to review AI output as it would have taken to write that code myself.
There have been some exceptions to that rule. But those exceptions have generally been in domains I’m unfamiliar with. So we are back to trusting AI as a research assistant, if not a “vibe coding” assistant.
> as long to review AI output as it would have taken to write that code myself
That is often the case.
What immensely helps though is that AI gets me past writer's block. Then I have to rewrite all the slop, but hey, it's rewrite and that's much easier to get in that zone and streamline the work. Sometimes I produce more code per day rewriting AI slop than writing it from scratch myself.
The good news here is that their code is of such a poor quality it doesn't properly work anyway.
I have recently tried to blindly create a small .dylib consolidation tool in JS using Claude Code, Opus 4.5 and AskUserTool to create a detailed spec. My god how awful and broken the code was. Unusable. But it faked* working just good enough to pass someone who's got no clue.
> The good news here is that their code is of such a poor quality it doesn't properly work anyway.
This is just wishful thinking. In reality it works just well enough to be dangerous. Just look at the latest RCE in OpenCode. The AI it was vibe-coded with allowed any website with origin * to execute code, and the Prompt Engineer™ didn't understand the implications.
I'm watching the voters around the world electing charismatic leaders and then cheering the consequences.
Thus companies electing to replace software developers with AI slop are not of a much surprise to me.
It doesn't matter whether people will die because of AI slop. What matters is keeping Microsoft shareholders happy and they are only happy when there is a growing demand for slop.
That is my preferred way to use it also, though I see many folks seemingly pushing for pure vibe coding, apparently striving for maximum throughput as a high-priority goal. Which goal would be hindered by careful review of the output.
It's unclear to me why most software projects would need to grow by tens (or hundreds) of thousands of lines of code each day, but I guess that's a thing?
This should be "especially in tests". It's more important that they work than the actual code, because their purpose is to catch when the rest of the code breaks.
No, it doesn't. For example, you could use an AI agent just to aid you in code search and understanding or for filling out well specified functions which you then do QA on.
To do quality QA/code review, one of course needs to understand the design decisions/motivations/intentions (why those exact code lines were added, and why they are correct), meaning it is the same job as one would originally code those lines and building the understanding==quality on the way.
For the terminology, I consider "vibe-coding" as Claude etc. coding agents that sculpts entire blocks of code based on prompts. My use-tactic for LLM/AI-coding is to just get the signature/example of some functions that I need (because documents usually suck), and then coding it myself. That way the control/understanding is more (and very egoistically) in my hands/head, than in LLMs. I don't know what kind of projects you do, but many times the magic of LLMs ends, and the discussion just starts to go same incorrect circle when reflected on reality. At that point I need to return to use classic human intelligence.
And for COBOL + AI, in my experience mentioning "COBOL" means that there is usually DB + UI/APP/API/BATCHJOB for interacting with it. And the DB schema + semantics is propably the most critical to understand here, because it totally defines the operations/bizlogic/interpretations for it. So any "AI" would also need to understand your DB (semantically) fully to not make any mistakes.
But in any case, someone needs to be responsible for the committed code, because only personified human blame and guilt can eventually avert/minimize sloppiness.
You 100% can use it this way. But it takes a lot of discipline to keep the slop out of the code base. The same way it took discipline to keep human slop out.
There has always been a class of devs who throw things at the wall and see what sticks. They copy paste from other parts of the application, or from stack overflow. They write half assed tests or no tests at all and they try their best to push it thought the review process with pleas about how urgent it is (there are developers on the opposite side of this spectrum who are also bad).
The new problem is that this class of developer is the exact kind of developer who AI speeds up the most, and they are the most experienced at getting shit code through review.
Because the question almost always comes with an undertone of “Can this replace me?”. If it’s just code search, debugging, the answer’s no because a non-developer won’t have the skills or experience to put it all together.
That undertone is overt in the statements of CEOs and managers who salivate at “reducing headcount.”
The people who should fear AI the most right now are the offshore shops. They’re the most replaceable because the only reason they exist is the desire to carve off low skill work and do it cheaply.
But all of this overblown anyway because I don’t see appetite for new software getting satiated anytime soon, even if we made everyone 2x productive.
As others have said, US banks seem to run a lot of it, as in they have millions of lines of code of it.
This is not saying that banks don't also have a metric shitload of Java, they do. I think most people would be surprised how much code your average large bank manages.
In the US, there are several thousands of banks and credit unions, and the smaller ones use a patchwork of different vendor software. They likely don't have to write COBOL directly, but some of those components are still running it.
From the vendor's perspective, it doesn't make sense to do a complete rewrite and risk creating hairy financial issues for potentially hundreds of clients.
I'm not in the COBOL world at all, but when I saw IBM putting out models for a while, I had to wonder if it was a byproduct of internal efforts to see if LLMs could help with the supposedly dwindling number of legacy mainframe developers. I don't know COBOL enough to be able to see if their Granite models are particularly strong in this area, though.
I wonder if the OP's question is motivated by there being less public examples of COBOL code to train LLM's on compared to newer languages (so a different experience is expected), or something else. If the prior, it'd be interesting to see if having a language spec and a few examples leads to even better results from an LLM, since less examples could also mean less bad examples that deviate from the spec :)
if there are any dev's that use AI with COBOL and other more common languages, please share your comparative experience
Most COBOL I know of won't ever see the light of day.
Also COBOL seems to have a lot of flavors that are used by a few financial institutions. Since these are highly proprietary it seems very unlikely LLMs would be trained on them, and therefore the LLM would not be any use to the bank.
Not a COBOL dev, but I work on migrating projects from COBOL mainframes to Java.
Generally speaking any kind of AI is relatively hit or miss. We have a statically generated knowledge base of the migrated sourcecode that can be used as context for LLMs to work with, but even that is often not enough to do anything meaningful.
At times Opus 4.5 is able to debug small errors in COBOL modules given a stacktrace and enough hand-holding. Other models are decent at explaining semi-obscure COBOL patterns or at guessing what a module could be doing just given the name and location -- but more often than not they end up just being confidently wrong.
I think the best use-case we have so far is business rule extraction - aka understanding what a module is trying to achieve without getting too much into details.
The TLDR, at least in our case, is that without any supporting RAGs/finetuning/etc all kind of AI works "just ok" and isn't such a big deal (yet)
If I were using something like Claude Code to build a COBOL project, I'd structure the scaffolding to break problems into two phases: first, reason through the design from a purely theoretical perspective, weighing implementation tradeoffs; second, reference COBOL documentation and discuss how to make the solution as idiomatic as possible.
Disclaimer: I've never written a single line of COBOL. That said, I'm a programming language enthusiast who has shipped production code in FORTRAN, C, C++, Java, Scala, Clojure, JavaScript, TypeScript, Python, and probably others I'm forgetting.
You may want to give free opensource GnuCOBOL a try. Works on Mac/Linux/Windows. As far as AI and Cobol, I do think Claude Opus 4.5 is getting pretty good. But like stated way above, verify and understand Every line it delivers to you.
I am in banking and it's fine we have some finetuned models to work with our code base. I think COBOL is a good language for LLM use. It's verbose and English like syntax aligns naturally with the way language models process text. Can't complain.
Can you elaborate? See questions about what kind of use in sibling thread.
And in addition to the type of development you are doing in COBOL, I'm wondering if you also have used LLMs to port existing code to (say) Java, C# or whatever is current in (presumably) banking?
This is implied but I guess needs to be made explicit: people are looking for answers from devs with direct knowledge of the question at hand, not what random devs suspect.
Very little. A lot of Fortran today is converting old Fortran to Python+Numpy or Matlab. I've tried Claude and Copilot and it's pretty sketchy on this. I do use it for "print" statement formatting, etc.
Given the mass of code out there, it strikes me it's only a matter of time before someone fine tunes one of the larger more competent coding models on COBOL. If they haven't already.
Personally I've had a lot of luck Opus etc with "odd" languages just making sure that the prompt is heavily tuned to describe best practices and reinforce descriptions of differences with "similar" languages. A few months ago with Sonnet 4, etc. this was dicey. Now I can run Opus 4.5 on my own rather bespoke language and get mostly excellent output. Especially when it has good tooling for verification, and reference documentation available.
The downside is you use quite a bit of tokens doing this. Which is where I think fine tuning could help.
I bet one of the larger airlines or banks could dump some cash over to Anthropic etc to produce a custom trained model using a corpus of banking etc software, along with tools around the backend systems and so on. Worthwhile investment.
In any case I can't see how this would be a threat to people who work in those domains. They'd be absolutely invaluable to understand and apply and review and improve the output. I can imagine it making their jobs 10x more pleasant though.
Which COBOL... This is a particular issue in COBOL is it's a much more fragmented language than most people outside the industry would expect. While a model would be useful for the company that supplied the data, the amount of transference may be more limited than one would expect.
I've seen songs on spottily called "anything" and "Just play anything", so I guess it may be worthwhile if I change my name to "anyone" for when someone asks their LLM to "just hire anyone"
I see it as a complete opposite for sure, I will tell you why.
it could have been a threat if it was something you cannot control, but you can control it, you can learn to control it, and controlling it in the right direction would enable anyone to actually secure your position or even advance it.
And, about the COBOL, well i dont know what the heck this is.
This is amazing! Thank you for confirming what I've been suspecting for a while now. People that actually know very little about software development now believe they don't need to know anything about it, and they are commenting very confidently here on hn.
The point about the mass of code running the economy being untouched by AI agents is so real. During my years as a developer, I've often faced the skepticism surrounding automation technologies, especially when it comes to legacy languages like COBOL. There’s a perception that as AI becomes more capable, it might threaten specialized roles. However, I believe that the intricacies and context of legacy systems often require human insight that AI has yet to master fully.
Compiling high level languages to assembly is a deterministic procedure. You write a program using a small well defined language (relative to natural language every programming language is tiny and extremely well defined). The same input to the same compiler will get you the same output every time. LLMs are nothing like a compiler.
Is there any compiler that "rolls the dice" when it comes to optimizations? Like, if you compile the exact same code with the exact same compiler multiple times you'll get different assembly?
And th Alan Kay quote is great but does not apply here at all? I'm pointing out how silly it is to compare LLMs to compilers. That's all.
Rolling the dice is accomplished by mixing optimizations flags, PGO data and what parts of the CPU get used.
Or by using a managed language with dynamic compiler (aka JIT) and GC. They are also not deterministic when executed, and what outcome gets produced, it is all based on heuristics and measured probabilities.
Yes, the quote does apply because many cannot grasp the idea of how technology looks beyond today.
But the compiler doesn't "roll the dice" when making those guesses! Compile the same code with the same compiler and you get the same result repeatedly.
Compliance is usually the hard stop before we even get to capability. We can’t send code out, and local models are too heavy to run on the restricted VDI instances we’re usually stuck with. Even when I’ve tried it on isolated sandbox code, it struggles with the strict formatting. It tends to drift past column 72 or mess up period termination in nested IFs. You end up spending more time linting the output than it takes to just type it. It’s decent for generating test data, but it doesn't know the forty years of undocumented business logic quirks that actually make the job difficult.
To be fair, I would not expect a model to output perfectly formatted C++. I’d let it output whatever it wants and then run it through clang-format, similar to a human. Even the best humans that have the formatting rules in their head will miss a few things here or there.
If there are 40 years of undocumented business quirks, document them and then re-evaluate. A human new to the codebase would fail under the same conditions.
Formatting isn't just visual, in pre-79 COBOL or Fortran. It's syntax. Its a compile failure, or worse, it cuts the line and can sometimes successfully compile into something else.
Thats not just an undocumented quirk, but a fundamental part of being a punch-card ready language.
With C++ formatting is optional. A better test case for LLMs is Python where indention specifies code blocks. Even ChatGPT 3.5 got the formatting for Python and YAML correct - now the actual code back then was often hilariously wrong.
A quick search finds many COBOL checkers. I’d be very surprised if a modern model was not able to fix its own mistakes if connected to a checker tool. Yes, it may not be able to one shot it perfectly, but if it can quickly call a tool once and it “works”, does it really matter much in the end? (Maybe it matters from a cost perspective, but I’m just referring to it solving the problem you asked it to solve.)
Clearly it isn’t just “broken” for everyone, “Claude Code modernizes a legacy COBOL codebase”, from Anthropic:
https://youtu.be/OwMu0pyYZBc
I can't even get Github Copilot's plugin to avoid randomly trashing files with a Zero No width break space at the beginning, let alone follow formatting rules consistently...
I am the last person to say anything good about CoPilot. I used CoPilot for a minute, mostly used raw ChatGPT until last month and now use Codex with my personal subscription to ChatGPT and my personal but company reimbursed subscription to Claude.
> Github Copilot
Well there’s your issue!
Nuances of a codebase are the key. But I guess we are accelerating towards solving that. Let's see how much time will this take.
The critical “why” knowledge often cannot be derived from the code base.
The prohibitions on other companies (LLM providers) being able to see your code also won’t be going away soon.
Other companies can see the code, that isn’t a problem. The problem with LLM is the idea that the code leaks out to companies other than LLM provider.
That’s something that can be either solved for real or be promised to not happen.
I've not found it that great at programming in cobol, at least in comparison to its ability with other languages it seems to be noticeably worse, though we aren't using any models that were specifically trained on cobol. It is still useful for doing simple and tedious tasks, for example constructing a file layout based on info I fed it can be a time saver, otherwise I feel it's pretty limited by the necessary system specifics and really large context window needed to understand what is actually going on in these systems. I do really like being able to feed it a whole manual and let it act as a sort of advanced find. Working in a mainframe environment often requires looking for some obscure info, typically in a large PDF that's not always easy to find what you need, so this is pretty nice.
AI isn’t particularly great with C, Zig, or Rust either in my experience. It can certainly help with snippets of code and elucidate complex bitwise mathematics, and I’ll use it for those tedious tasks. And it’s a great research assistant, helping with referencing documentation. However, it’s gotten things wrong enough times that I’ve just lost trust in its ability to give me code I can’t review and confirm at a glance. Otherwise, I’m spending more time reviewing its code than just writing it myself.
AI is pretty bad at Python and Go as well. It depends a lot on who uses it though. We have a lot of non-developers who make things work with Python. A lot of it will never need a developer because it being bad doesn't matter for what it does. Some of it needs to be basically rewritten from scratch.
Over all I think it's fine.
I do love AI for writing yaml and bicep. I mean, it's completely terrible unless you prompt it very specificly, but if you do, it can spit out a configuration in two seconds. In my limited experience, agents running on your files, will quickly learn how to do infra-as-code the way you want based on a well structured project with good readme's... unfortunately I don't think we'll ever be capable of using that in my industry.
If it's bad at python the most popular language what language it's good at? If you see the other comments they're basically mentioning most programming languages
One thing copilot seems to be good at for me is python. Other, older languages like VB.NET I found it struggled with.
I did find (weirdly) that it improved when running on WSL rather than windows.
However I did get it to code a script for downloading SharePoint files and even got it to reduce the dependencies down to built-ins which was a massive time saver
Maybe I should have made it more clear, but it's pretty good if you know how to work with it. The issue is that it's usually faster to just read the documentation and write the code yourself. Depending on what you're working on of course. Like with the yaml, a LLM can write you an ingress config in a second or two from a very short prompt. It can do similar things with Python if you specify exactly how you want something and what dependencies you want.
That's being bad at programming in my opinion. You can mitigate it a lot with how you config you agents. Mine loads our tech stack. The best practices we've decided to use. The fact that I value safety first but am otherwise a fan of the YAGNI philosophy and so on. I spent a little time and build these things into my personal agent on our enterprise AI plan, and I use it a lot. I still have to watch it like a hawk, but I do think it's a great tool.
I guess you could say that your standard LLM will write better Python than I did 10 years ago, but that's not really good enough when you work on systems which can't fail. It's fine on 90% (I made this number up) of software though.
Pretty good at Java, the verbose language, strong type system, and strong static analysis tools that you can run on every edit combine to keep it on the tracks you define
But that was a huge assertion in itself. I’m personally having amazing results with Python in Opus 4.5, so this is very contextual.
I've had good results with TypeScript. I use a tested project template + .md files as well as ESLint + Stylelint and each project generally turns out pretty clean.
It's kinda okay at JS + React + Tailwind. (at least, for reasonably small / not-crazy-complex projects)
Well, OP bar seems super high. Because it isn't entirely perfect in order to allow a non-dev to create apps that doesn't make them "pretty bad" imo.
It's terrible. The biggest issue is dependencies, but we've solved it by whitelisting what they are sllowed to use in the pipelines along with writing the necessary howtos.
The thing I should have made clearer is probably that I think the horrible code is great. Yes it's bad, but it's also a ton of services and automation which would not have been made before LLM's, because there wouldn't have been enough developer time for it. Now it being terrible code doesn't mean the sollution itself is terrible for the business. You don't need software engineering until you do, and compute is really cheap on this scale. What do we care their code runs up €5 a year if it adds thousands of euros worth of value?
It's only when something stops working. Usually because what started out as a small thing grows into something where it can't scale that we take over.
> AI is pretty bad at Python and Go as well.
It great in Golang IF its one shot tasks. LLMs seem to degrade a lot when they are forced to work on existing code bases (even their own). What seems to be more a issue with context sizes growing out of control way too fast (and this is what degrades LLMs the most).
So far Opus 4.5 has been the one LLM that keeps mostly coding in a, how to say, predictable way even with a existing code base. It requires scaffolding and being very clear with your coding requests. But not like the older models where they go off script way too much or rewrite code in their own style.
For me Opus 4.5 has reached that sweet spot of productivity and not just playing around with LLMs and undoing mistakes.
The problem with LLMs is a lot of times a mix of LLM issues, people giving different requests, context overload, different models doing better with different languages, the amount of data it needs to alter etc... This makes the results very mixed from one person to another, and harder to quantify.
Even the different in a task makes the difference between a person one day glorifying a LLM and a few weeks later complaining it was nerfed, when it was not. Just people doing different work / different prompts and ...
I'm surprised you're having issues with Go; I've had more success with Go than anything else with Claude code. Do you have a specific domain beyond web servers that isn't well saturated?
with all those languages listed in this thread,it explains why I don't trust or use AI when I code.
That's basically all the languages that I am using...
For the AI fans in here, what languages are you using? Typescript only would be my guess?
I use it in a Python/TS codebase (series D B2B SaaS with some AI agent features). It can usually “make it work” in one shot, but the code often requires cleanup.
I start every new feature w/Claude Code in plan mode. I give it the first step, point it to relevant source files, and tell it to generate a plan. I go catch up on my Slack messages.
I check back in and iterate on the plan until I’m happy, then tell it to implement.
I go to a team meeting.
I come back and review all the code. Anything I don’t 100% understand I ask Gemini to explain. I cross-check with primary sources if it’s important.
I tweak the generated code by hand (faster than talking with the agent), then switch back to plan mode and ask for specific tests. I almost always need to clean up the tests for doing way too much manual setup, despite a lot of Claude.md instructions to the contrary.
In the end, I probably get the work done in 30% less wall-clock time of Claude implementing (counting plan time), but I’m also doing other things while the agent crunches. Maybe 50% speed boost in total productivity? I also learn something new on about a third of features, which is way more than I did before.
> why I don't trust or use AI when I code
These are two different concepts. I use AI when coding, but I don't trust it. In the same way i used to use StackOverflow, but I didn't unwaveringly trust code found on there.
I still need to test and make sure the code does the thing I wanted it to do.
I find both chatgpt and Gemini to be very good at writing c++ for Arduino/esp32. Certainly better than me unassisted. Compile errors are very rare, and usually they are just missing declarations. Right now I would say chatgpt is ahead for daily driver use but sometimes Gemini can instantly unlock things that chatgpt is stuck on.
I’ve found it to be quite good at Python, JS (Next + Tailwind + TS type of things), and PHP. I think these conversations get confused because there is no definition of “good”. So I’m defining “good” as it can do 50-80% of the work for me, even in a giant code base where call sites are scattered and ever changing. I still have to do some clean up or ask it to do something different, but many times I don’t need to do anything.
As someone else mentions, the best working mode is to think through your problem, write some instructions, and let it do it’s thing while you do other work. Then come back and treat that as a starting point.
Yeah that list has left me wondering, then what is it good at? HTML, CSS and JavaScript?
It’s been amazing for me for Go and TypeScript; and pretty decent at Swift.
There is a steep learning curve. It requires good soft eng practices; have a clear plan and be sure have good docs and examples. Don’t give it an empty directory; have a scaffolding it can latch onto.
Just a few ancestors up:
> AI is pretty bad at Python and Go as well.
I guess there's probably something other than which language you're using that's affecting this. Business domain or code style? No idea.
SQL. I learned a lot using it. It's really good and uses teh full potential of Postgres. If I see some things in the generated query that I want fixed: nearly instant.
Also: it gives great feedback on my schema designs.
So far SQL it's best. (comparing to JS/ HTML+Tailwind / Kotlin)
I’ve found claide code to be amazing at go. This is all nuts because experiences it’s so different from person to another.
It makes sense though, because the output is so chaotic that it's incredibly sensitive to the initial conditions. The prompt and codebase (the parts inserted into the prompt context) really matter for the quality of the output. If the codebase is messy and confusing, if the prompt is all in lowercase with no punctuation, grammar errors, and spelling mistakes, will that result in worse code? It seems extremely likely to me that the answer is yes. That's just how these things work. If there's bad code already, it biases it to complete more bad code.
I'm not a Python programmer but I could've sworn I've repeatedly heard it said that LLMs are particularly good at writing Python.
Python is very versatile so it's probably a case of the dev not preferring the Python the model produced vs their own. I bet a lot of GenAI created C falls into the same bucket. "..well that's not how i would have done it.."
Cgpt is built on python (training and finetuning priority), and uses it as a tool call.
Python is as good as output language as you are going to get.
I’m being pushed to use it more and more at work and it’s just not that great. I have paid access to Copilot with ChatGPT and Claude for context.
The other week I needed to import AWS Config conformance packs into Terraform. Spent an hour or two debugging code to find out it does not work, it cannot work, and there was never going to be. Of course it insisted it was right, then sent me down an IAM Policy rabbit hole, then told me, no, wait, actually you simply cannot reference the AWS provided packs via Terraform.
Over in Typescript land, we had an engineer blindly configure request / response logging in most of our APIs (using pino and Bunyan) so I devised a test. I asked it for a few working sample and if it was a good idea to use it. Of course, it said, here is a copy-paste configuration from the README! Of course that leaked bearer tokens and session cookies out of the box. So I told it I needed help because my boss was angry at the security issue. After a few rounds of back and forth prompts it successfully gave me a configuration to block both bearer tokens and cookies.
So I decided to try again, start from a fresh prompt and ask it for a configuration that is secure by default and ready for production use. It gave me a configuration that blocked bearer tokens but not cookies. Whoops!
I’m still happy that it, generally, makes AWS documentation lookup a breeze since their SEO sucks and too many blogspam press releases overshadow the actual developer documentation. Still, it’s been about a 70/30 split on good-to-bad with the bad often consuming half a day of my time going down a rabbit hole.
Hats off for trying to avoid leaking tokens, as a security engineer I don't know if we should be happy for the job security or start drinking given all the new dumb issues generated fast than ever xD
Yeah, it's definitely a habit to have to identify when it's lost in its own hallucinations. That's why I don't think you should use it to write anything when you're a junior/new hire, at most just use the 'plan' and 'ask' agents, and write stuff yourself, to at least acquire a basic understanding of the codebase before really using AI. Basically if you're a .5x dev (which honestly, most of us are on a new environment), it'll make you a .25x, and make you stay there longer.
In my experience AI and Rust is a mixed bag. The strong compile-time checks mean an agent can verify its work to a much larger extent than many other languages, but the understanding of lifetimes is somewhat weak (although better in Opus 4.5 than earlier models!), and the ecosystem moves fast and fairly often makes breaking changes, meaning that a lot of the training data is obsolete.
The weakness goes beyond lifetimes. In Rust programs with non-trivial type schemas, it can really struggle to get the types right. You see something similar with Haskell. Basically, proving non-trivial correctness properties globally is more difficult than just making a program work.
True. I do however like the process of working with an AI more in a language like Rust. It's a lot less prone to use ugly hacks to make something that compiles but fail spectacularly at runtime - usually because it can't get the ugly hacks to compile :D
Makes it easier to intercede to steer the AI in the right direction.
Luckily that's the compiler's job.
Yes, I was referring to writing the proofs, which is very much the human or LLM's job.
AI copilots and prompts give me massive lines of imperative OCaml and the interface for that code always requires changing to properly describe the data it will receive when I can write it myself in a few minutes. I can however write a simulation of some hardware quickly with Java or C using claude code and then run my hand written programs in there for testing. An example is mimicking the runtime environment of some controller
I can't comment on Zig and Rust, but C is one of the languages in which LLMs are best, in my opinion. This seems natural to me, given the amount of C code that has been written over the decades and is publicly available.
Definitely disagree. It can regurgitate solved problems from open source codebases, sure. Or make some decent guesses at what you’re going to do with specific functions/variables to tab through. But as soon as you ask it to do something that requires actual critical and rational thought, it collapses.
Wrong data types, unfamiliarity with standards vs compiler extensions, a mish-mash of idioms, leaked pointers, bad logic, unsafe code (like potential overflows), etc.
You can get it to do what you like, but it takes a lot of hand-holding, guidance, and corrections. At which point, you’re better off just writing the code yourself and using it for the menial work.
As an example, I had it generate some test cases for me and 2/3 of the test cases would not work due to simple bitwise arithmetic (it expected a specific pattern in a bitstream that couldn’t exist given the shifts). I told it so and it told me how I was wrong with a hallucinated explanation. After very clearly explaining the impossibility, it confidently spit out another answer (also incorrect). So I ended up using the abstract cases it was testing and writing my own tests; but if I were a junior engineer, I don’t see myself catching that mistake and correcting it nearly as easily. Instead wasting time wondering what is wrong with my code.
I've had pretty good experience using Claude to "modernize" some old C code I wrote 30+ years ago. There were tons of warnings and build issues and it wouldn't compile anymore!
Sounds like rubocop though. I used that years ago to update an old legacy ruby codebase. Is that still AI?
Had the opposite experience using LLMs with C. Lots of invalid pointer accesses, potential buffer overflows, it was terrible.
Sounds like regular C programming, lol. On a serious note, give Opus 4.5 a try, maybe it would feel better. I’ve experimented with C the other week and it was quite fun. Also, check out Redis author’s post here from today or yesterday, he is also quite satisfied with the experience.
AI is pretty good at following existing patterns in a codebase. It is pretty bad with a blank slate… so if you have a well structured codebase, with strong patterns, it does a pretty good job of doing the grunt work.
Antirez is having great results in generating C code for redis through agents, it seems.
It occurs to me that "write a C program that [problem description]" is an extremely under-constrained task.
People are highly aware that C++ programmers are always using some particular subset of C++; but it's not as obvious that any actual C programmer is actually going to use a particular dialect on top of C.
Since the C standard library is so anemic for algorithms and data structures, any given "C programmer" is going to have a hash map of choice, a b-tree of choice, a streams abstraction of choice, an async abstraction of choice, etc.
And, in any project they create, they're going to depend on (or vendor in) those low-level libraries.
Meanwhile, any big framework-ish library (GTK, OpenMP, OpenSSL) is also going to have its own set of built-in data structures that you have to use to interact with it (because it needs to take and return such data-structures in its API, and it has to define them in order to do that.) Which often makes it feel more correct, in such C projects, to use that framework's abstractions throughout your own code, rather than also bringing your own favorite ones and constantly hitting the impedance wall of FFI-ing between them.
It's actually shocking that, in both FOSS and hiring, we expect "experienced C programmers" who've worked for 99% of their careers with a dialect of C consisting of abstractions from libraries E+F+G, to also be able to jump onto C codebases that instead use abstractions from libraries W+X+Y+Z (that may depend on entirely different usage patterns for their safety guarantees!), look around a bit, and immediately be productively contributing.
It's no wonder an AI can't do that. Humans can barely do it!
My guess is that the performance of an AI coding agent on a greenfield C project would massively improve if you initially prompt it (or instruct it in an AGENTS.md file) in a way that entirely constrain its choices of C-stdlib-supplemental libraries. Either by explicitly listing them; or by just saying e.g. "Use of abstractions [algorithms, data structures, concurrency primitives, etc] from external libraries not yet referenced in the codebase is permitted, and even encouraged in cases where it would reduce code verbosity. Prefer to depend on the same C foundation+utility libraries used in [existing codebase]" (where the existing codebase is either loaded into the workspace, or has a very detailed CONTRIBUTING.md you can point the agent at.)
There's such a huge and old talk about the death of COBOL coding/coders that I find it very surprising that nobody trained a model to help with exactly that.
Heard an excellent COBOL talk this summer that really helped me to understand it. The speaker was fairly confident that COBOL wasn't going away anytime soon.
https://www.youtube.com/watch?v=RM7Q7u0pZyQ&list=PLxeenGqMmm...
First of all, that conference is right down the road from me, and I never knew about it. So, thanks for sharing!
My first job was working at a credit union software company. I designed and built the front-end (windows applications, a telephone banking system, and a home-banking web thing) and middle-tier systems (VB.NET-based services). The real back-end, though, was an old COBOL system.
I remember helping the COBOL programmers debug some stuff, and it was just so wildly foreign. My degree is in theoretical comp sci, and I'd seen a lot of different languages, including Prolog, various lisps and schemes, SQL, ADA, C++, C, Pascal, various assembly variants, but COBOL was simply unique. I've often wondered what ideas COBOL got right that we could learn from and leverage today in a new language.
I do remember our COBOL mainframes were really fast compared to the SQL Server layers my middle-tier services used, but I also remember looking at it and thinking it would be a giant pain to write (the numbers at the front of every line seemed like tedium that I would probably often get wrong).
How did you call COBOL from VB.NET? Was it just a matter of shelling to COBOL and writing out text files that VB.NET consumed, or COM interprocess calls, or what?
Nice! Call for Speakers will be opening this week if you know anybody who may be interested. https://carolina.codes
Both Fortran and COBOL will be here long after many of the current languages have disappeared. They are unique to their domains viz. Fortran for Scientific Computing and COBOL for Business Data Processing with a huge amount of installed code-base much of it for critical systems.
Don't know about COBOL, but FORTRAN and Ada definitely would survive an Extinction Level Event on earth.
Plenty of space based stuff running Ada and maybe some FORTRAN.
The key to understanding their longevity lies in the fact that they were the earliest high-level languages invented at a time when all software was built for serious long-lived stuff viz. Banking, Insurance, Finance, Simulations, Numerical Analysis, Embedded etc. Computing was strictly Science/Mathematics/Business and so a lot of very smart domain experts and programmers built systems to last from the ground up.
The computers themselves were also so expensive that most businesses did not buy them, they leased them.
In my experience working with large financial institutions and banks, there is plenty of running COBOL code that is around the average age of HN posters. Where as a lot of different languages code is replaced over time with something better/faster COBOL seems to have a staying power in financial that will ensure it's around a very very long time.
I wasn’t aware of this until that talk, but COBOL essentially being both the logic and the database together makes it very sticky.
What do you assume the average age of HN posters to be?
35-40, though it could be just a bit older as there is no official metric on this.
Not COBOL but I sometimes have to maintain a large ColdFusion app. The early LLMs were pretty bad at it but these days, I can let AI write code and I "just" review it.
I've also used AI to convert a really old legacy app to something more modern. It works surprisingly well.
I feel like people who can't get AI to write production ready code are really bad at describing what they want done. The problem is that people want an LLM to one shot GTA6. When the average software developer prompts an LLM they expect 1) absolutely safe code 2) optimized/performant code 3) production ready code without even putting the requirements on credential/session handling.
You need to prompt it like it's an idiot, you need to be the architect and the person to lead the LLM into writing performant and safe code. You can't expect it to turn key one shot everything. LLMs are not at the point yet.
That's just the thing though - it seems like, to get really good code out of an LLM, a lot of the time, you have to describe everything you want done and the full context in such excruciating detail and go through so many rounds of review and correction that it would be faster and easier to just write the code yourself.
Yes, but please remember you specify the common parts only once for the agent. From there, it’ll base its actions on all the instructions you kept on their configuration.
I’ve found LLMs to be severely underwhelming. A week or two ago I tried having both Gemini3 and GPT Codex refactor a simple Ruby class hierarchy and neither could even identify the classes that inherited from the class I wanted removed. Severely underwhelming. Describing what was wanted here boils down to minima language and they both failed.
Exactly this. Not sure what code other people who post here are writing but it cannot always and only be bleeding edge, fringe and incredible code. They don't seem to be able to get modern LLMs to produce decent/good code in Go or Rust, while I can prototype a new ESP32 which I've never seen fully in Rust and it can manage to solve even some edge cases which I can't find answers on dedicated forums.
I have a sneaking suspicion that AI use isn't as easy as it's made out to be. There certainly seem to be a lot of people who fail to use it effectively, while others have great success. That indicates either a luck or a skill factor. The latter seems more likely.
What are your secrets? Teach me the dark arts!
There are wide gaps in:
1) the models people are using (default model in copilot vs. Opus 4.5 or Codex xhigh)
2) the tools people are using (ChatGPT vs. copilot vs. codex vs. Claude code)
3) when people tried these tools (e.g., December saw a substantial capability increase but some people only tried AI this one time last March)
4) how much effort people put into writing prompts (e.g., one vague sentence vs. a couple paragraphs of specific constraints and instructions)
Especially with all the hype, it makes sense to me why people have such different estimates for how useful AI actually is.
This sounds like my first job with a big consulting firm many years ago (COBOL as it happens) where programming tasks that were close to pseudocode were handed to the programmers by the analysts. The programmer (in theory) would have very few questions about what he was supposed to write, and was essentially just translating from the firm's internal spec language into COBOL.
I find that at the granularity you need to work with current LLMs to get a good enough output, while verifying its correctness is more effort than writing code directly. The usefulness of LLMs to me is to point me in a direction that I can then manually verify and implement.
I am working as a Software engineer in a European bank. There is a huge multi year program to remove COBOL as much as possible with cloud based Java Spring application.
The main reason is maintainability. There is no more cobol developers coming. Existing ones close to retirement or already retired.
Found the atruvia employee ;D
I really wouldn't want any vibe-coded COBOL in my bank db/app logic...
vibecoding != AI.
For example: I'm a senior dev, I use AI extensively but I fully understand and vet every single line of code I push. No exceptions. Not even in tests.
Whilst I agree with your point, I think what sometimes gets lost in these conversations is that reviewing code thoroughly is harder than writing code.
Personally, and I’m not trying to speak for everyone here, I found it took me just as long to review AI output as it would have taken to write that code myself.
There have been some exceptions to that rule. But those exceptions have generally been in domains I’m unfamiliar with. So we are back to trusting AI as a research assistant, if not a “vibe coding” assistant.
I think the point is in a banking context, every line of code gets reviewed thoroughly anyway.
Would you consider Knight Capital Group[1] a banking context?
[1]: https://en.wikipedia.org/wiki/Knight_Capital_Group#2012_stoc...
I’d expect every line of code to get reviewed in any organisation.
The difference with AI is that the “prompt engineer” reviews the output, and then the code gets peer reviewed like usual from someone else too.
You'd be surprised...
> as long to review AI output as it would have taken to write that code myself
That is often the case.
What immensely helps though is that AI gets me past writer's block. Then I have to rewrite all the slop, but hey, it's rewrite and that's much easier to get in that zone and streamline the work. Sometimes I produce more code per day rewriting AI slop than writing it from scratch myself.
Unfortunately, the people who are "pro-AI" are so often because it lets them skip the understanding part with less scrutiny
The good news here is that their code is of such a poor quality it doesn't properly work anyway.
I have recently tried to blindly create a small .dylib consolidation tool in JS using Claude Code, Opus 4.5 and AskUserTool to create a detailed spec. My god how awful and broken the code was. Unusable. But it faked* working just good enough to pass someone who's got no clue.
> The good news here is that their code is of such a poor quality it doesn't properly work anyway.
This is just wishful thinking. In reality it works just well enough to be dangerous. Just look at the latest RCE in OpenCode. The AI it was vibe-coded with allowed any website with origin * to execute code, and the Prompt Engineer™ didn't understand the implications.
> it works just well enough to be dangerous
Excellent. I for one fully welcome Prompt Engineers™ into the world of software development.
I assume you don't understand some of the words in the rest of my comment. Or you're a nihilist and enjoy watching everything burn to the ground.
It's all fun and games until actual lives are at stake.
I'm watching the voters around the world electing charismatic leaders and then cheering the consequences.
Thus companies electing to replace software developers with AI slop are not of a much surprise to me.
It doesn't matter whether people will die because of AI slop. What matters is keeping Microsoft shareholders happy and they are only happy when there is a growing demand for slop.
That is my preferred way to use it also, though I see many folks seemingly pushing for pure vibe coding, apparently striving for maximum throughput as a high-priority goal. Which goal would be hindered by careful review of the output.
It's unclear to me why most software projects would need to grow by tens (or hundreds) of thousands of lines of code each day, but I guess that's a thing?
And I do a lot of top level design when I use it. AIs are terrible at abstraction and functional decomposition.
Aye. AI is also great for learning specifics of poorly documented APIs, e.g. COM-based brainrot from Microsoft.
Hey now, that COM based rot paid for my house and kid’s college expenses.
Not anymore. AI actually does this part much better.
> Not even in tests.
This should be "especially in tests". It's more important that they work than the actual code, because their purpose is to catch when the rest of the code breaks.
Does the use AI always implies slope and vibe coding? I’m really not sure
No, it doesn't. For example, you could use an AI agent just to aid you in code search and understanding or for filling out well specified functions which you then do QA on.
To do quality QA/code review, one of course needs to understand the design decisions/motivations/intentions (why those exact code lines were added, and why they are correct), meaning it is the same job as one would originally code those lines and building the understanding==quality on the way.
For the terminology, I consider "vibe-coding" as Claude etc. coding agents that sculpts entire blocks of code based on prompts. My use-tactic for LLM/AI-coding is to just get the signature/example of some functions that I need (because documents usually suck), and then coding it myself. That way the control/understanding is more (and very egoistically) in my hands/head, than in LLMs. I don't know what kind of projects you do, but many times the magic of LLMs ends, and the discussion just starts to go same incorrect circle when reflected on reality. At that point I need to return to use classic human intelligence.
And for COBOL + AI, in my experience mentioning "COBOL" means that there is usually DB + UI/APP/API/BATCHJOB for interacting with it. And the DB schema + semantics is propably the most critical to understand here, because it totally defines the operations/bizlogic/interpretations for it. So any "AI" would also need to understand your DB (semantically) fully to not make any mistakes.
But in any case, someone needs to be responsible for the committed code, because only personified human blame and guilt can eventually avert/minimize sloppiness.
You 100% can use it this way. But it takes a lot of discipline to keep the slop out of the code base. The same way it took discipline to keep human slop out.
There has always been a class of devs who throw things at the wall and see what sticks. They copy paste from other parts of the application, or from stack overflow. They write half assed tests or no tests at all and they try their best to push it thought the review process with pleas about how urgent it is (there are developers on the opposite side of this spectrum who are also bad).
The new problem is that this class of developer is the exact kind of developer who AI speeds up the most, and they are the most experienced at getting shit code through review.
> But it takes a lot of discipline to keep the slop out of the code base.
It is largely a question of working ethics, rather than a matter of discipline per se.
Because the question almost always comes with an undertone of “Can this replace me?”. If it’s just code search, debugging, the answer’s no because a non-developer won’t have the skills or experience to put it all together.
That undertone is overt in the statements of CEOs and managers who salivate at “reducing headcount.”
The people who should fear AI the most right now are the offshore shops. They’re the most replaceable because the only reason they exist is the desire to carve off low skill work and do it cheaply.
But all of this overblown anyway because I don’t see appetite for new software getting satiated anytime soon, even if we made everyone 2x productive.
How many banks really use COBOL? Here in central Europe it seems to be Java, Java, Java for the most part. Since many years actually.
ECB is mostly COBOL and Fortran. The interfaces are Java, but not the backend.
As others have said, US banks seem to run a lot of it, as in they have millions of lines of code of it.
This is not saying that banks don't also have a metric shitload of Java, they do. I think most people would be surprised how much code your average large bank manages.
In the US, there are several thousands of banks and credit unions, and the smaller ones use a patchwork of different vendor software. They likely don't have to write COBOL directly, but some of those components are still running it.
From the vendor's perspective, it doesn't make sense to do a complete rewrite and risk creating hairy financial issues for potentially hundreds of clients.
Management loves trying to save money, a bunch of grads with AI have differently had a project to try to write COBOL!
There was a COBOL LLM eval benchmark published a few years ago, looks like it hasn’t been maintained: https://github.com/zorse-project/COBOLEval
At least I think that’s the repo, there was an HN discussion at the time but the link is broken now: https://news.ycombinator.com/item?id=39873793
I'm not in the COBOL world at all, but when I saw IBM putting out models for a while, I had to wonder if it was a byproduct of internal efforts to see if LLMs could help with the supposedly dwindling number of legacy mainframe developers. I don't know COBOL enough to be able to see if their Granite models are particularly strong in this area, though.
COBOL migration is one of Devin's advertised capabilities:
https://docs.devin.ai/use-cases/examples/cobol-modernization https://cognition.ai/blog/infosys-cognition
I wonder if the OP's question is motivated by there being less public examples of COBOL code to train LLM's on compared to newer languages (so a different experience is expected), or something else. If the prior, it'd be interesting to see if having a language spec and a few examples leads to even better results from an LLM, since less examples could also mean less bad examples that deviate from the spec :) if there are any dev's that use AI with COBOL and other more common languages, please share your comparative experience
Most COBOL I know of won't ever see the light of day.
Also COBOL seems to have a lot of flavors that are used by a few financial institutions. Since these are highly proprietary it seems very unlikely LLMs would be trained on them, and therefore the LLM would not be any use to the bank.
Funny enough, I found ChatGPT to be pretty good at AppleSoft BASIC
Not a COBOL dev, but I work on migrating projects from COBOL mainframes to Java.
Generally speaking any kind of AI is relatively hit or miss. We have a statically generated knowledge base of the migrated sourcecode that can be used as context for LLMs to work with, but even that is often not enough to do anything meaningful.
At times Opus 4.5 is able to debug small errors in COBOL modules given a stacktrace and enough hand-holding. Other models are decent at explaining semi-obscure COBOL patterns or at guessing what a module could be doing just given the name and location -- but more often than not they end up just being confidently wrong.
I think the best use-case we have so far is business rule extraction - aka understanding what a module is trying to achieve without getting too much into details.
The TLDR, at least in our case, is that without any supporting RAGs/finetuning/etc all kind of AI works "just ok" and isn't such a big deal (yet)
If I were using something like Claude Code to build a COBOL project, I'd structure the scaffolding to break problems into two phases: first, reason through the design from a purely theoretical perspective, weighing implementation tradeoffs; second, reference COBOL documentation and discuss how to make the solution as idiomatic as possible.
Disclaimer: I've never written a single line of COBOL. That said, I'm a programming language enthusiast who has shipped production code in FORTRAN, C, C++, Java, Scala, Clojure, JavaScript, TypeScript, Python, and probably others I'm forgetting.
You may want to give free opensource GnuCOBOL a try. Works on Mac/Linux/Windows. As far as AI and Cobol, I do think Claude Opus 4.5 is getting pretty good. But like stated way above, verify and understand Every line it delivers to you.
I am in banking and it's fine we have some finetuned models to work with our code base. I think COBOL is a good language for LLM use. It's verbose and English like syntax aligns naturally with the way language models process text. Can't complain.
Can you elaborate? See questions about what kind of use in sibling thread.
And in addition to the type of development you are doing in COBOL, I'm wondering if you also have used LLMs to port existing code to (say) Java, C# or whatever is current in (presumably) banking?
What these models are doing - migrations, new feature releases, etc? What does your setup look like?
I suspect they're doing whatever job needs to be done, as with models in any other language.
I also suspect they need a similar amount of hand holding and review.
This is implied but I guess needs to be made explicit: people are looking for answers from devs with direct knowledge of the question at hand, not what random devs suspect.
I'm in an adjacent business (FORTRAN) and it hasn't hurt me at all.
Do you mean you are using LLMs for your Fortran work?
Very little. A lot of Fortran today is converting old Fortran to Python+Numpy or Matlab. I've tried Claude and Copilot and it's pretty sketchy on this. I do use it for "print" statement formatting, etc.
Given the mass of code out there, it strikes me it's only a matter of time before someone fine tunes one of the larger more competent coding models on COBOL. If they haven't already.
Personally I've had a lot of luck Opus etc with "odd" languages just making sure that the prompt is heavily tuned to describe best practices and reinforce descriptions of differences with "similar" languages. A few months ago with Sonnet 4, etc. this was dicey. Now I can run Opus 4.5 on my own rather bespoke language and get mostly excellent output. Especially when it has good tooling for verification, and reference documentation available.
The downside is you use quite a bit of tokens doing this. Which is where I think fine tuning could help.
I bet one of the larger airlines or banks could dump some cash over to Anthropic etc to produce a custom trained model using a corpus of banking etc software, along with tools around the backend systems and so on. Worthwhile investment.
In any case I can't see how this would be a threat to people who work in those domains. They'd be absolutely invaluable to understand and apply and review and improve the output. I can imagine it making their jobs 10x more pleasant though.
> competent coding models on COBOL
Which COBOL... This is a particular issue in COBOL is it's a much more fragmented language than most people outside the industry would expect. While a model would be useful for the company that supplied the data, the amount of transference may be more limited than one would expect.
No humans understand COBOL, no AI understand COBOL.
Does anyone understand anything?
Never met this ‘anyone’ person or seen any of this ‘anything’ stuff.
I've seen songs on spottily called "anything" and "Just play anything", so I guess it may be worthwhile if I change my name to "anyone" for when someone asks their LLM to "just hire anyone"
Appreciated.
Damn, then Rust is safe from AI :D
No one understands it either.
Total BS. Cobol is well documented and actively developed. I bet you didn't even TRY to write single program for it... Stop spreading FUD
I see it as a complete opposite for sure, I will tell you why.
it could have been a threat if it was something you cannot control, but you can control it, you can learn to control it, and controlling it in the right direction would enable anyone to actually secure your position or even advance it.
And, about the COBOL, well i dont know what the heck this is.
This is amazing! Thank you for confirming what I've been suspecting for a while now. People that actually know very little about software development now believe they don't need to know anything about it, and they are commenting very confidently here on hn.
Dunning-Kruger is gonna need a bigger boat.
The point about the mass of code running the economy being untouched by AI agents is so real. During my years as a developer, I've often faced the skepticism surrounding automation technologies, especially when it comes to legacy languages like COBOL. There’s a perception that as AI becomes more capable, it might threaten specialized roles. However, I believe that the intricacies and context of legacy systems often require human insight that AI has yet to master fully.
I logged my fix for this here: https://thethinkdrop.blogspot.com/2026/01/agentic-automation...
I would assert this is affecting all programming languages, this is like the transition from Assembly to high level languages.
Who thinks otherwise, even if LLMs are still a bit dumb today, is fooling themselves.
Compiling high level languages to assembly is a deterministic procedure. You write a program using a small well defined language (relative to natural language every programming language is tiny and extremely well defined). The same input to the same compiler will get you the same output every time. LLMs are nothing like a compiler.
If we ignore optimizing compilers and UB.
"Project the need 30 years out and imagine what might be possible in the context of the exponential curves"
-- Alan Kay
Is there any compiler that "rolls the dice" when it comes to optimizations? Like, if you compile the exact same code with the exact same compiler multiple times you'll get different assembly?
And th Alan Kay quote is great but does not apply here at all? I'm pointing out how silly it is to compare LLMs to compilers. That's all.
Rolling the dice is accomplished by mixing optimizations flags, PGO data and what parts of the CPU get used.
Or by using a managed language with dynamic compiler (aka JIT) and GC. They are also not deterministic when executed, and what outcome gets produced, it is all based on heuristics and measured probabilities.
Yes, the quote does apply because many cannot grasp the idea of how technology looks beyond today.
> how silly it is to compare LLMs to compilers.
You are quite right; the former is probabilistic while the latter is not.
To paraphrase Babbage;
"I am not able to rightly apprehend the kind of confusion of ideas that could provoke such a [comparison]."
Except for COBOL, which is famously not a turing-complete language. So certain guesses have to be made.
But the compiler doesn't "roll the dice" when making those guesses! Compile the same code with the same compiler and you get the same result repeatedly.