| > that's so much hype not grounded in anything. Not grounded in anything? Really, anything? There's no evidence you've seen in the last three years that has had any impact on what you think is possible with AI coding? I'm not a game dev by trade, I am AI researcher turned AI engineer (not the "AI engineer" of the modern era, like getting ML deployed in prod). That being said, I have tried and failed to make plenty of games over the years. Every mistake in the book. Making my own engine, trying to do 3D game as first approach, making mobile apps 2010, using Unity for personal projects, switching to Godot. I'm not at your level in game dev, but I've seen plenty at scale. > I'm guessing your job is to promote this stuff based on that being your purpose of going to GDC I'm not an AI hype maker, I'm a stunned observer. A person who has said "no way" AI will impact coding for the last 10 years. I am watching evidence in front of my very eyes and my workflow and simply adapting to the reality on the ground along with reasonable projection of capabilities into the future (based on velocity and reasonable assumptions). I don't need to convince you of anything, but I'm confident your beliefs won't hold up against the stockpile of evidence gathering now and in the near future. Since you took a swing and miss on who I am, I'm going to guess that you are reluctant to incorporate these tools into your own workflow (perhaps having cast them off quickly after some frustrated probing) and have a deep-seated skepticism of the tech that doesn't really matter what I say. Where's my inference coming from? The idea that you've conflated that AI coding by its very nature needs to be buggy and unstable. I wish you luck on holding the line as it slips past us on an exponential curve. |
I actually do use them (daily), and I do find them to be useful. I even bought a couple books about how to implement them, and my own current app has an MCP server I've built. I also find them to be overhyped and not particularly essential (if I had to choose one to take to a desert island, I would take a good IDE over a good AI, for example). The way I currently see them is they make things more convenient, but it doesn't really expand what's possible for me, and that convenience can come at a profound cost (some of which we're just now starting to see). I already enjoy coding, so the convenience of not having to code is not that big of a win for me, and when you factor in the extra factors I don't even think it makes me go that much faster. (I either have to correct its outputs, or spend extra time understanding them, or the codebases become unwieldy and I've DEFINITELY seen that. On average AI code seems to be about 2-3x more verbose than what I would write). I can see why someone that can't code (or doesn't want to) would find them to be amazing, but I also think those people are not in a great position to judge the outputs, which is still pretty necessary and probably will be for the foreseeable future.
I think a common thread with AI use is it seems amazing when you're asking it questions about fields you're not an expert in (I like to ask ChatGPT questions about biology for instance), but it seems very much not amazing when you ask it questions about fields you actually are an expert in.
> Where's my inference coming from? The idea that you've conflated that AI coding by its very nature needs to be buggy and unstable. I wish you luck on holding the line as it slips past us on an exponential curve.
See, but that's what bothers me about these conversations in general, people just go "look at number go up!" like you can take any trend and just keep growing it forever. Or people make really bold claims with no evidence like your "you can vibecode an engine in 6 months" claim. It's hard to even _learn_ all the features of a mature game engine like Unreal Engine over _multiple_ years, and to do so would require at least being partially fluent in multiple different very different domains (art, animation, level design, shaders, workflows, about a trillion 3rd party tools, etc.) How could you vibe code something that quickly when even experts probably know about 15-25% of the feature set? Like how can you prompt an AI to build features you don't even understand? Also these engines evolved over a long period of time, so there are so many micro decisions and customer feedback loops that are completely opaque but important and are embedded in the code. The code in many engines contains decades of important knowledge!
I'm not swayed by the "exponentials" argument because past results do not guarantee future performance. I'm basing my opinions on what I've seen in my job and my personal work, and what I've seen is they're a boost, but not to the degree being sold and I'm kind of sick of people telling me that I need to be on board with the hype train or get "left behind". AI is not that hard to prompt, I don't view that as a real skillset, so to me I can either continue to bet on my own skills and intellect and hope they still hold value, or I can just get out of the industry entirely because my role would be entirely pointless. I would rather bet on myself even if I'm wrong, because all becoming dependent on it is going to do is degrade my skillset and my intellect, and if I'm right, well, honestly it's not particularly hard to train up on current AI harnesses or whatever compared to learning a skill (my computer science degree took 4 years -- learning Claude took a day. In what universe is knowing how to use Claude a marketable skill?)