[Reticularas]

No one in game and game engine development is using ChatGPT.

It doesn't even work, even if it was productive, as barely any problems the average gamedev faces on a day-to-day have been surfaced to the training data of an LLM.

[softfalcon]

Yeah, this exactly. The only person I know who is succeeding at "game dev" with ChatGPT is an older bloke I know who just began their baby steps on a pet project with Unity.

Sure, if you're still learning how to move position from x to x+1, ChatGPT can help you with those basics, but anything more than that and they start hitting me up on Discord looking for real answers.

Something as simple as how the serialization of fields in behaviour classes was too much for ChatGPT. They read what it hallucinated up, got confused, and were calling me up to explain minutes later. For those who don't know Unity, this is something you have to do daily when building new logic into game components and ChatGPT was clueless about it.

The limitations of ChatGPT as a programming aid are very very obvious and it has a long way to go before it's really useful to seasoned professionals.

[Rapzid]

ChatGPT 4 may be better but CoPilot is terrible at making suggestions on algorithm work. It'll slip a -(x) in to parameters where it's supposed to be positive. Use the sim but wrong variable names. Swap parameters around. It's can be a nightmare and cost you way more time debugging than it saves typing.

Specifically with type-ahead suggestions I believe with typing you are more engaged in thinking about what is correct. With a CoPilot suggestion the tendency is to just skim it to see if it look good. And that's probably part of the problem; the LLM is trained on what looks good!

This may look good:

  computeNodeDistance(sourceNodesA, destNodesA, x, -1)

But it's supposed to be:

  computeNodeDistance(sourceNodesA, destNodesB, x, 1)

Good times.

[xgkickt]

Not until legal approve it, which doesn't seem like any time soon.