I’ve rewrote the post to be “more pleasant” to read. Sorry for the unhinged original post.
I’ve a devlog on the card game thing I’m working on. This is a side project, a thing that I’m using for learning and practing some of the stuff from the AI perspective.
This is my discovery and other things, all feedback, ideas and concepts are welcome on the perspective of the software and programming.
Thanks in advance
Very sorry to hear that. Would you mind elaborating on the topic? What was the issue? The language? The tone? The way that I described? English isn’t my mother language, and I want to improve how I do express myself in it.
I’ll start with the unnecessarily combative tone with your reader. Then the self-contradicting absolutes (claude sucks, but my solution is x% as good as the lowest-end claude). Then the fact that it’s really just an untargeted rant against “everything” that doesn’t really justify the anger.
And you hit my pet peeve about saying “containers aren’t running on bare metal” when they run “on bare metal” exactly as much as non-containerized processes. That’s the point of containers.
Edit: I did rewrote the post. Once I calmed my mind down I think it’s better now. At least my closed friends prefer this version of it.
That was really insightful.
And align with a close friend on the topic. So you know. I think I’ll rewrite this post in a less “aggressive” tone.
When two of my close friends don’t like a post, there is a good chance that the post isn’t good enough to leave the oven.
That’s a much nicer read! I’m interested in some of your Claude experiences though.
I tested Claude heavily, and I have to be direct, it consistently fell below my quality bar. Even with strict, detailed guidelines, the code it produced was riddled with subtle issues, off-by-one errors, improper error handling, and a complete disregard for my idiomatic Go patterns.
Which Claude model were you using? Opus 3.7 & 3.8 tend to do a very good job at things like that. It catches my off-by-one issues much better than I do… They are, however, a bit slower as you point out. But my local models on an NVidia 3070 with only 8Gig of VRAM are garbage…
Single-letter variable names might be fine for tight loops, but when you’re generating thousands of lines of code via AI, i, j, and k become impossible to trace.
I find that if I drop in a CLAUDE.MD some style guides that it does a good job of following them. In particular I hate the Python idiom of prepending “_” to “private” variables and functions. If I put a note in there it stops it from doing it.
As an aside - i, j, and k for loop variables is a very old tradition going back to FORTRAN where i, j, and k were always integer variables. So i is the outer loop, j the inner loop, and k the inner-inner loop. If you have more nested loops than that you’re doing it wrong…
But if you don’t like it the style hint should help.
Most of my complains are on the func (f *thing) name(w *http.Writer...), this is ok for small functions, but when you have 20 lines, where 10 of then uses single letters around are unbearable.
Regarding the CLAUDE.md, didn’t work at all. I don’t know if the training material is so small in the GoLang that makes the model go crazy on specific details, or maybe the single letter was preferred since it consumes less tokens, either way, didn’t work at all. And I was using the Opus. I know that they released the Fable, but I can’t use since it’s coding related.
The Qwen model I used a quantized version of it for a lot of the stuff, it’s super fast, but a 8 vRAM won’t hold it. On my Linux box I have 16vRAM and managed to hold it tight, but I notice that I was destroying the GPU for long running tasks, so I aborted.
I do have a Blackwell in mind, with some luck I can manage to grab one and test a full precision DeepSeek V4 Flash in it. I already saw some people doing it in a single Blackwell, so should be a thing to have at home to play around, but this is a far future.
Regarding the CLAUDE.md, didn’t work at all.
Not at all? I do know that LLMs tend to have a “recency bias” and lose earlier context over time, especially when Claude summarizes a session to create a new one seamlessly. I’ve run into issues with it wanting to run git commands despite me explicitly telling it that it should never do so… Keeping sessions short seems to help.
For my use case, didn’t work. But I was in a more spec driven than anything else. So maybe with a shorter spec file would have a better result.
