cross-posted from: https://hexbear.net/post/8882542
It’s a different story for the more established studios with an existing following and previous titles. Game Oracle found that the use of AI by these studios resulted in a significant 40% to 60% drop in sales.
That’s a huge difference. AI stigma seems to hit competent developers with a lot to lose the hardest, and I’m not sure that game studios are ready to accept it.
Every coworker I’ve seen that uses AI code tools heavily is bad. They produce (or at least push) nonsense code they don’t understand.
I would rather have a team that goes slowly and understands what they’re building than a team of excitable slop pushers going a thousand lines a second.
I think that a lot of the issue behind AI is profit motive. If that was removed from the equation, AI could be a really useful tool for development. But the fact that the main usecase at every company is “go faster so line goes up” gets in the way of that.
The other issue is efficiency. Burning down a rainforest to make a stupid banking app isn’t very appealing to me. But if we could work to make generative AI more efficient at what it does, and transition to consumer ownership of compute/hardware, it might be more feasible. That’s a long ways off, and I don’t think the powers that be want that future.
I use AI every day at work. We built an entire orchestration framework on top of Claude Code. It features skills that can pull in a jira ticket with well defined acceptance criteria and complete it without very much human intervention at all. We’ve built out entire epics and our QA team has not seen an uptick in defects. This is because we all still do manual code reviews in addition to AI reviews. We even have a skill that checks AC against the PR diff to make sure we’re meeting AC before it gets to QA. AI does make our team more efficient.
But at what cost? None of us are learning anything about coding. I feel more burnt out than I ever have. It’s an environment nightmare. It is a moral/ethical nightmare. Finally, none of us are seeing any additional compensation for improving efficiency. That compensation is going to Anthropic.
There is a world where AI can be a net benefit to the world, but it isn’t our current one.
TL;DR I basically agree lol
I understand what you’re saying, but you’re talking about real world examples rather than the mechanics at play.
Let’s said you have the best engineer in the world. They are fantastic as describing nuanced, complex ideas. The fastest they can write is about 300 words per minute. The fastest they can read is 1200 words per minute.
Put them up against an AI model. They write at 6000 wpm & read at like 11 mil or something ridiculous.
Now, you’re making the argument, “Speed isn’t everything!” and that’s true. Which would rather you have, though: the same engineer, the AI, or the same engineer using the AI? I’d argue you’d rather have them both, because you have someone who can describe what they want in depth, they can point it towards heuristics or targets, and they can setup evals or governance strategies to better control the output.
I’m not trying to be dismissive, but I work with competent, smart people. My experiences have been the opposite. 🤷
How are you talking about the mechanics at play when the price of tokens isn’t a part of your discussion?
There’s many facets to it, including token costs, the risk of captured markets, more “win more” economics.
Right now, this stuff is insanely cheap as a solo dev who isn’t making a bunch of autonomous agent API calls. Nearly all my interactions are working with an agent to make a game or project. I’m not attempting to, nor interested in, yet another AI service based company.
I work at a reasonable pace, usually 1 conversation is my focus with actual work, whereas another agent may be doing research or documentation. Too many things in motion turns into a mess to test or validate. Solve one issue at a time. I don’t hit token limits, and I’m making good progress.
When I read stories of people being inefficient or wasteful, yeah, that’s sad, don’t do that. I feel reasonably efficient & the cost is really low (for me/lifestyle). I use the $20/mon Claude subscription for my personal.
Local models have no token prices.
The problem has never ever ever been words per minute. That is a completely irrelevant metric. A distraction.
Anything the AI produces is going to need to be evaluated by a person, and that is a more difficult, less rewarding task.
And if it doesn’t need to be reviewed by a person because it’s magically flawless, that’s extremely anti-labor so fuck that.
It’s harder to review code than to write code. On our team reviewing has always been the bottleneck. Faster output would actually make things harder in some cases.
Good luck. They’ve been taught to hate AI, and their brains shut off as soon as you say anything positive about it.