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
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
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.
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.
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
Agree. It’s excellent at python all round. If it lays out things how you want it to is a matter of preference and usually requires prompting it to restructure. That’s the standard way you work with AI code gen though, it’s iterative and requires testing. If you do it well it can be specified up front as a style guide set of instructions
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 ...
> 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.
I find this to be true only if you have very explicit rules in CLAUDE.md and even then it still messes up.
I have "you will use the shared code <here>" twice in my CLAUDE.md as it will constantly write duplicate code.
Something that is also annoying is that if it moves some code somewhere with the intent to slightly modify it I've seen it delete the code, then implement from scratch, and then modify it to what it has been specified to do. This completely breaks tests. I then have to say "look at this earlier commit - you've caused a complete regression."
This is a workflow boundary problem showing up as a tool problem.
When changes aren’t constrained by explicit inputs and checkpoints, models optimise locally and regress globally.
Predictability comes from the workflow, not the model.
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?
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 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’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.
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.
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)
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.
How is this an issue specifically with Rust and Haskell? Do you find that LLMs have an easier time proving global correctness with C, Python, or Typescript?
Do you have examples of LLMs proving global correctness for say, C? Having worked on static analysis for both C and Rust, Rust is the easier problem because of the type system, but I am eager to be proven wrong!
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 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
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.)
"De facto" is the keyword there. Only the nation of origin has any say on company management and infrastructure in a de jure manner. The only power non-origin nations/entities have is via leveraging their ability to do business in the region and/or their local holdings.
The dependencies they're likely referring to aren't core libraries, they're shared interfaces. If you're using protobufs, for instance, and you share the interfaces in a repo. Updating Service A's interface(s) necessitates all services dependent on communicating with it to be updated as well (whether you utilize those changes or not). Generally for larger systems, but smaller/scrappier teams, a true dependency management tree for something like this is out of scope so they just redeploy everything in a domain.
> If you're using protobufs, for instance, and you share the interfaces in a repo. Updating Service A's interface(s) necessitates all services dependent on communicating with it to be updated as well (whether you utilize those changes or not).
This is not true! This is one of the core strengths of protobuf. Non-destructive protobuf changes, such as adding new API methods or new fields, do not require clients to update. On the server-side you do need to handle the case when clients don't send you the new data--plus deal with the annoying "was this int64 actually set to 0 or is it just using the default?" problem--but as a whole you can absolutely independently update a protobuf, implement it on the server, and existing clients can keep on calling and be totally fine.
Now, that doesn't mean you can go crazy, as doing things like deleting fields, changing field numbering or renaming APIs will break clients, but this is just the reality of building distributed systems.
What you are talking about is simply keeping the API (whether a library or a service) backwards-compatible. There are plenty strategies to achieve that, and it can be done with almost any interface layer (HTTP, protobuf, JSON, SQL, ...).
I was oversimplifying for the sake of example, but yes you are correct. Properly managed protobufs don't require an update on strict interface expansion; so shouldn't always require a redeploy.
A demake would be a reimagining of a modern game into the style and aesthetics of the time. E.g. taking God of War and turning it into a 2D Shinobi-style platformer for Sega Genesis. Or turning Gran Turismo into a Mode7-style racer on SNES.
In this case, the creator wrote a custom 3D renderer and recreated the models/meshes to get as close of an approximation of the N64 experience onto the GBA.
I wouldn't call it a port necessarily ("recreation" seems more apt), but it's closer to that than a demake.
The game was built for the Saturn and ported to PS1 (versus being built independently for it). During that process, graphical fidelity and effects were lost and audio transcoded in a lossy manner. Sprites were lower quality, the water effects were worse, textures were downgraded, shadows were lost, etc.
This is a pretty bad demonstration since they took literally zero effort in syncing the two side-by-side til the end, but gives you an idea:
That being said, it's not so much worse that PS1 is unplayable. It's just that if you had the choice between the two, the Saturn version is slightly more pleasant to look at and listen to.
Do you not know that the US is a Federal system and there are (at minimum) 50 different ways that schools are funded?
California's schools (for instance) aren't funded by local taxes, they're funded by the state and allocated funding based on a formula[1] of performance, need, population, etc. They can be augmented by local taxes, but in practice that's rare as the wealthy just avoid the system altogether; instead, opting for private institutions.
That's at least 12% of the population that is not funded in the manner you outline.
Equity remains a valid criticism of LCFF in California specifically.
For one unremarkable observation in this area, see the following think tank report:
> States often commission cost studies to establish the level of funding required to help students meet state standards. LPI analyzed five of the more recent of these studies [...] All of these studies recommended additional weighted funding to support English learners and students considered "at-risk," which was most often defined by a measure of family income and also included other factors [...] The recommended weights for English learners in these studies ranged from 15% to 40% of the base grant level in each state. The recommended weights for at-risk students ranged from 30% to 81%. Compared to the recommended funding in these states, the LCFF’s supplemental grant weight of 20% is at the lower end of the recommended range of weights for English learners and below the range of weights for at-risk students.
Win98 was head and shoulders above System 9, from a stability perspective. It had protected memory, actual preemptive multitasking, a somewhat functional driver system built on top of an actual HAL, functional networking, etc, etc.
To be clear, Win98 was a garbage fire of an OS (when it came to stability); which makes it so much worse that Mac OS 8-9 were so bad.
98's multitasking and memory 'protection' were a joke. In the same mid high machine for the era, 2k and xp were miles ahead of w98 on mid-high load.
Maybe not on a Pentium, but once you hit 192MB of RAM and some 500 MHz P3/AMD k7, NT based OSes were tons better.
You only realized that upon opening a dozen of IE windows. W98, even SE, will sweat. 2k will fly.
On single tasks, such as near realtime multimedia ones, w98 would be better, such as emulators/games or video players with single thread decoders. On multiprocessing/threading, w98 crawled against W2K even under p4's and 256MB of RAM.
Well, Win NT is an actual operating system, and Win 98 and Classic macOS are just horribly overgrown home computer shells in an environment they should never have been exposed to.
Ahem, w98 BSOD if you sneezed hard near it. Installing a driver? BSOD. IE page fault? BSOD. 128k stack limit reached? either grind to a halt or a BSOD. And so on...
I worked at a company that was delivering a client-side app in Java launched from IE. I think we had an ActiveX plugin as the "launcher." This predated "Java Web Start." It was hysterically bad. We were targeting 32 meg Win 98 systems and they were comically slow and unstable. Most of our developers had 64 and 128 meg boxes with NT 4.0. I mostly worked on the server side stuff, and used them as a terminal into the Solaris and Linux systems.
reply