By using this site, you agree to our Privacy Policy and our Terms of Use. Close
Zkuq 2 days ago
Edit Thread
IcaroRibeiro said:
Zkuq said:

I support AI in gaming too, without any limitations set in stone either - in theory. In practice, my main concern is how the training material for AI has been picked with little regard for ownership/copyrigh, and last I heard, that was still legally a grey area at best. As for quality, the end result will speak for itself (for better or worse). I suspect for many aspects, AI is perfectly adequate - like do I really care if some of the most widely used textures in a game, e.g. many ground textures, are created by humans? No, I don't think I do. I do think quality is still often an issue with AI, but in many (not all!) cases, it's also probably good enough that most of us will never pay attention to it unless going out of our way just to look for it.

As for jobs or art, I don't think AI is the issue. For all I care, AI can replace all the jobs, and it's up to the society to figure out how people are going to make a living and how they can be happy. If we think some kind of an AI tax is a good solution, I'm all for it. If it's something else, that's fine too. The same goes for art: the society should support the creation of art to the extend it's deemed valuable (and to be clear, I think it is valuable). If AI disrupts the society, the society should adapt. Regardless, if AI brings efficiency benefits (which is probably questionable in many fields), AI will come, or we will fall behind the likes of China, because the likes of China will adapt AI and they will find ways to keep the population happy enough. Adapt or survive is what we have to do, assuming AI doesn't turn out to be a bubble (which it still might).

This... doesn't match my experience at all. AI is definitely useful for programming but it's because of everything else it can do besides actual coding that it's useful for: figuring things out, explaining them, test automation... But for actual coding, it still does really stupid mistakes and tends to ignore existing coding conventions from the project. For things like games, stupid mistakes in particular might be something that are acceptable, but where I work, that's simply not acceptable. The code AI produces is also often overly verbose, which actually makes it harder to understand.

Also, as far as I know, AI has generally been trained with most code that's available, not just best, so it kind of tends to produce 'average' code. Of course that's still better than most developers seem to be capable of doing (at least based on my limited experience).

Regarding the training using all code (including shit code), that are metrics applied during the training phase to qualify what is the output of the model. Even if there is lots of trash code in training phase the learning algorithm will strongly punish low quality answers (in this case, shit code) and give good scores to high quality answers (high quality code)

I strongly recommend using the Premium version of Claude. It has ability to understand contextual project patterns and generate new code following the same principles, even if you need to explicitly tells it to follow the coding standards already present in the source code. The level of actual coding erros is minimal for almost everything outside front-end development strongly related to HTML rendering and user interaction (among other more niche and esoteric applications)

For automation, backend development, machine learning and anything strongly logic-based it's much more efficient, fast and produces much less bugs than humans 

I'm using Gemini 3 in my current project (mostly data engineering code) and it's absurd how you can quickly write shit code to teach Gemini what bussines logic you actually want and it will change your slop to bug-free production-ready code instantly. Team has being using it 5 months and the number of bugs related to actual coding were reduced to literal zero. Unit testing has been rendered completely useless, because the AI already cover edge cases and side effects

All bugs now are related to badly defined (or badly understood) bussines logic, or problems on input data. Really game changer 

We're working with a relatively large codebase, much of which could be called legacy and much of that legacy we're still building on because we still have a lot of other tech debt that's prioritized higher for fixing but still need to further develop the system, and AI is simply unreliable. It's not necessarily bad, but it can't understand the codebase the way humans can, so it makes dumb mistakes because of that, but from time to time it also makes dumb mistakes unrelated to that. I'm sure it would do better in a smaller project or a project that was built to higher standards from the ground up, but unfortunately it's not.

For what we're doing, the risk of any bugs resulting from poor AI code are simply such that basically everything the AI does needs to be reviewed, preferably carefully, lest we risk stupid mistakes getting through. It's slow work because of that, maybe still somewhat faster than doing it manually but reviewing so much code is not fun.

We're using paid versions of several models, I believe, and that includes Claude. I think the models available to us are top-notch, for what is available at the moment at least.

I'm glad it's working out for you, but for us, there's a long way to go still before AI can really be a game-changer. For us specifically, I feel like it's more like a minor convenience at best at the moment.