游戏开发：Durante 谈游戏开发中 AI 的思考（来自一线）[Reddit]

One thing that I think few might realise is that games are, at least for now, actually less directly affected by AI than other software development tasks. Obviously the focus here is on discussing those effects on games, and as a consequence it really might not seem like this is the case. But it's really not too complicated. The primary reason is that agent systems thrive when there is (1) a fast feedback loop, and (2) a clear objective function to measure.

In many fields in software development, those are actually easier to establish than in games. Regarding (1), especially real-time games really are not something current agent systems and harnesses are particularly well-equipped to deal with. It can of course be done, but at tremendous cost and low speed compared to e.g. observing the outcome of a web query, measuring the efficiency of a simulation, or verifying a business process on a database. And regarding (2), that is even worse, since e.g. something like "good-feeling gameplay", which is often also a function of code, is really hard to quantify.

This also means that some types of development in the gaming space -- e.g. ports, clones, and reverse engineering -- are inherently more suitable for current agent systems and will make faster progress at the same cost. And I think we clearly see this in practice right now.

As a side note, I think this feedback loop / objective function issue is actually why, despite being large language models, I personally think they are actually more effective in (particular parts of) math and programming than "simply" writing prose. For both, you can measure and prove at least some objective qualities, and iterate on them. That's not generally the case for writing.