While messing around with ChatGPT, I was excited to learn that i could program in eZ80 as I intended to learn. To test its abilities, I gave it the simple task of a Tic-Tac-Toe game. After over 30 minutes of debugging and 20 code rewrites, it still was unable to make a functioning game. Might use it in the future for documentation help but it does not have the range I had hoped.

P.S. I can include the code if anyone is interested.

ChatGPT is useless at programming pretty much anything.

Also important to note that the less common of a language it is, the less it'll know about it and something like ez80 assembly isn't something that's used a lot compared to other languages like C or Python. Definitely better to ask your questions either here in the forums or in SAX, there's usually someone around who could probably help or at least point you in the right direction.

Yeah, I was mostly just playing around with ChatGPT for fun. I wanted to know its capabilities for coding, specifically for eZ80.



I intend to browse resources and lurk on the forum a lot during the summer. If I can't find answers, I'll definitely reach out to you as opposed to the bot. Thanks for being such a kind and supportive community!

I believe that Chat GPT is a great assistant to have when coding. After all it is going to replace simple and repetitive jobs in the future. But I believe when it comes to coding in programming a languages that doesn't really have a variety of open source programs before 2021 it tents to struggle due to the lack of data used to train the model.


To a certain extent. To date, Chat GPT "3.0" is roughly 80% correct, with the remaining 20% being less valuable languages that were left out of train data. However, if Chat GPT receives more training data, I predict that it's overall 80% correctness will be closer to a 85% - 90%. (Urhm Chat GPT "4.0" does this)

This further solidifies my belief that you are an idiot.

Hey MateoConLechuga, Simply going off assumptions doesn't guarantee you correctness. you simply don't have any experience with technology (NW "Neural Networks", LLM "Large Language Models") and you probably using the free version of ChatGPT. Where as the paid version has slightly fixed most of the issues you are addressing.

What's wild is people arguing over AI. It is absolutely something that you can use. Yes ez80 asm is lacking in the AI department, but you have to "teach" it. Work with it. It's a tool. It's like people talk trash about AI, but go ahead, compile your code into .8xp without the compiler every time you make a code change. Do it by hand. Oh, can't avoid using tools? I'm pretty sure historically people have argued over tools being released and used widespread but ultimately it's you who does something with it. Idk why people complain about tools when you literally use tools to do everything there is to do around development. Ultimately it's all tools designed by people who put work into them, whether IDEs, compilers and now AIs. It's all to help you progress and it should be left at that.

One recommended thing I'd say is to try is to make it write it in c and compile it. Then use debugger to grab the machine asm code. It may or may not be able to rewrite it with the right back and forth feedback and persuasion. It's like teaching a child. It's a tool and shouldn't replace anyone, but if you're writing something and genuinely don't have time and nobody else has done it, nobody else has a say so as to if you can't do it. If they won't do it who will eh?

Yes, I've also had issues with ez80 asm and AI. But AI does improve. It couldn't write things it can write now, because we're all training it.

I love arguing about AI! Let me jump in!



That is true, AI is absolutely something you can use. No one can stop you. The question is, should you?



This is an interesting argument for sure. It may not be perfect, but you have to teach it. Is that really characteristic of tools, though? I've never once had to teach a hammer how to hammer in a nail, or for that matter a compiler how to compile code. I just grab them and they work. That convenience is, in my opinion, their main appeal. I would be far less likely to use a compiler that didn't just work off the shelf!



Hm... okay... well, first off, if I didn't have a compiler, honestly? I probably wouldn't bother writing in C at all, but just use assembly. If you don't let me have an assembler at all, then I would be stuck hand-assembling, which people did for a good long time and it basically worked.

But all of that is missing the point. Yes, I really would rather not avoid using a compiler or an assembler. And yet, I can perfectly well avoid using an LLM. LLMs are not an important part of my workflow. You seem to want to make this distinction impossible by calling them both "tools," but one seems far more useful to me than the other, so I don't really see why lumping them together like this makes sense.



Again, compilers and IDEs (well, actually, I don't use an IDE, but whatever) just work, whereas AI, as you point out, doesn't. It sounds like you're the one putting in the work on the AI front, given all your talk about "teaching" it.



And I guess here's where I really get into the weeks: AI doesn't help me, and I'm not convinced it helps you, either.

You're the Minecraft guy, right? I've seen that development thread, and I can't help but notice that it petered out somewhat. Now admittedly, I don't always finish my projects (Elite, for example, I just couldn't really be bothered to finish bug-testing, because it would take soooo long...), but I suspect that the disappearance of that thread from the firehose might be down to you hitting a bit of a limit to what you're able to squeeze out of ChatGPT. And even if that isn't the case, I contend that using AI both reduces your productivity and hinders your learning.

A study earlier this year found that AI makes developers, on average, 19% slower, even if they predicted that they would be, on average, 24% faster using AI than not. And anecdotally, pretty much every experienced developer I know agrees that the time spent arguing with the bot, in the end, isn't worth it compared to just doing the task yourself. Not to mention the lower quality of LLM-generated code, which (by definition) doesn't have the same thought put into it as human-written code.

I don't want to make too many assumptions, but I don't think that you are a super experienced programmer. Not that there is anything wrong with that-- I think it's true of most users on this site. There are a lot of high schoolers on here, and for a lot of people, this community is a place for learning programming in an interesting and kind of niche setting. If you fall into this group in any way, I am very happy that you are here, and interested in a hobby that we share!

That said, I truly believe that LLMs are devastating to learning to program. Instead of learning how to actually think about the program space, design algorithms, and write code, you are primarily learning how to deal with the fickle outputs of a chat box. For as you are "teaching" it, it is training you to be dependent on it, and the more time and energy you sink into AI, the more invested you become. You already act like it's impossible for you to even conceptualize of throwing it away-- it's as though you were being asked to discard a compiler, you say. But is that really the kind of dependence that should be encouraged, especially among novice developers?



This prescription makes no sense. First off, that's already output by the compiler as an intermediate step, assuming that you're using the same toolchain as the rest of us; I think it's ld.asm? Somewhere in your obj/ folder, you will find that, so there is no reason to get CEmu involved.

Second, calling that the "assembly source" for your program is patently absurd. You are getting none of the benefits that you would have by writing in assembly, and just bloating the file size and decreasing the maintainability of just using the true C source.



OP: If you don't have time to write something by hand, you don't have time to vibe-code it. Again, there is no real time savings. And Timmy, what's all this about replacing anybody?



If they won't... what, use AI? It's not like anybody has to do anything mentioned in this thread.



This is, indisputably, true, which is a great note to leave off on. But I want to append a more cautious qualification: even if you, Timmy, are pitching in some more ez80 training data (as are others, though not always with their consent), that volume is nothing compared to what exists for more common languages like Python or technologies like C for standard desktop operating environments. I think it would take a very long time indeed for the amount of training data featuring ez80 on the CE and friends to be sufficient for LLMs to have anywhere near the broad range of basic competence that they seem to display for more mainstream use-cases. So while I admire your spirit, I have to say that your project seems, on its face, rather unlikely to succeed.

AI is alright at general coding tasks, but for anything requiring even a modicum of complexity, you better make sure you scrutinize what it gives you or don't use it at all because it makes mistakes. Lots of them.
There are some models that are better trained for coding than ChatGPT. But even they...get things wrong. I have a friend at work who works with data analytics like LLMs and even ones extensively trained for coding..need to be babysat.

Use them at your own risk.

I find it hilarious that you can tell ChatGPT to write a program and ask it if it has any bugs. It will say no. Then, you can feed its own program back into it and tell it to debug it and it will catch all of its own mistakes.
Anyway, I think ChatGPT is pretty well developed for Python and can certainly learn programming languages. But take the code with the same skepticism you would if it came from your amateur programming friend, if not with more. Unlike your friend, most LLMs will give you programs with the false confidence that they're prefect.

Why trust AI with eZ80 ASM when it can barely write in TI-BASIC https://www.cemetech.net/forum/viewtopic.php?t=20539

I do appreciate the feedback, good to see you agree and disagree on specific parts. My points were mainly from my own experience. I use Cursor. It constantly gets updated and I see improvement in it the more I use it. On my GitHub, I’ve used it to port a simple-ish vsmile emulator from windows (written in C/cpp) to Android. I had it create everything, and compiled in android studio and tested. The 3D Minecraft vibe coded was only stopped because it was intended as AI testing only for me, and it indeed could not write ez80 asm code. The closest I could get is to get the asm code from the compiler output and try getting the AI to optimize it, which it couldn’t really do. It also failed every time to link asm into C code. That’s where my bad experiences end.
Otherwise it’s basically did things for me I don’t have time to do. I coded a lot in high school (because during school I didn’t pay attention and instead coded). I learned a bit and can code myself, but not optimized at all and don’t claim to know everything. I now have a full time job and it’s mandatory overtime, it takes away from my free time and in the little free time I have I don’t code. I only get realistically 3-4 hours at most of my own time on workdays. Weekends I could code but don’t.
Now what I will say is, AI does help me personally. It’s created a battery saver GUI app for me on my gaming laptop running Linux, I have gained a bit extra time now. It’s ported repo (steam game) to android. I gave it android flash images and it flashed lineageos to my device, even modified super.img for me. It’s written the 3D Minecraft project and I even threw my old calculator projects at it. It split my messy one-file C code into separate .c and .h and worked out the compiler errors. I also give it pc error codes or tell it to read my event viewer logs itself in windows to work out crashes, blue screen or application crashes. With Cursor, I can let it access any file on my pc and it can work with the terminal, run commands and read output. That’s what I mean with training it and working with it, also that it’s a tool.
I do it for things that I use personally. Usually to do things I want done and don’t want to spend more time doing it.

This topic was about ez80 asm coding, and so I spoke from experience, I couldn’t get anything to work. It kept trying to do z80 asm I think, but anything else so far I’ve done personally it’s worked well enough for my satisfaction.

ChatGPT might not really have any data fed to it about TI Basic either

It also sucks when trying to do even less-known languages, like CASIO BASIC. It completely fails at that, according to what I have seen.

its especially bad since its not connected to the internet. i would try grok ai its worked for me somewhat.
also grok ai is somewhat ok at ez80/z80

ChatGPT is ok for text based games especially in Python but it is terrible for TI Basic and any form of assembly on the TI 84 and similar. Also ChatGPT is VERY slow when writing code

Thats because these languages are not known as well as other languages

I actually made a pretty cool text based RPG in Python using ChatGPT, but it really only works for the basic frame of the game. I still had to write a decent bit of it myself using Pycharm. Still no graphics though. I can see the appeal for the programmers who do not have enough time to code the basic outline of the game.

[D] dsteffen said this on the sax chat and I think this is relevant.
I've been using Gemini as a second set of eyes on my code. It caught a nasty bug in an unrolled ldi loop I had. I had loaded b register for djnz, but forgot all about c getting decremented along the way. doh! Pasted the code into Gemini and it caught it in 10 seconds