A small team spent six years on Big Walk, and this week the conversation is about a fraction of a percent of its code. I wrote about why House House didn't owe anyone an apology, and why punishing honest AI disclosure teaches developers to say less. https://t.co/57061wH9ey
@LaurenceIpsum People made games just fine before Unity and Unreal, too. Nobody says you’re not a real game developer because you didn't write your own custom renderer. Better tools still leave you with plenty of problems to solve, and that's where the experience and ingenuity come from now.
I think creators should get to decide whether AI has a place in their process. If you don't want to use it, that’s your choice. People who do should be able to use it too. 🤷♂️
@FatBoyBreeder@turtlekiosk I agree developers shouldn't lie. Steam didn't require disclosure of this coding assistance. IGF suggested it, and House House disclosed it. They got attacked anyway. They shouldn't have to apologize for their choice of tools. That only discourages honesty.
@onlyarchangelI He’s the co-founder and president of Valve. They’re the company behind Steam and games like Deadlock, Team Fortress 2, Half-Life, and Portal.
House House says it wrote the disclosure because it was a suggested part of its IGF submission. Steam doesn't ask about code or chatbot use. I want developers to feel able to answer openly, and it's sad to see the studio apologizing after doing that.
I agree that jurors can vote by their own values. But I disagree with asking IGF to disallow games "containing any AI-generated work." I don't see how a festival could check that from a finished game. The rule would mostly catch the developers who report their process honestly.
I find it sad that House House is apologizing for this. The studio estimates that well under 0.1% of Big Walk's code involved LLM use. In its example, a chatbot suggested a possible fix for a Unity shadow problem an engineer at Unity had concluded couldn't be solved. The team had to change the suggestion extensively, and the fix was about 10 lines of code.
As far as I can tell, the studio's honest disclosure is why Big Walk became the center of this debate. I respect the jurors' right to vote by their own values, but I disagree with making that festival policy for everyone.
This neural network plays Breakout perfectly, and it was never trained. 🧵
Fable 5.2 read the game's source code and wrote all 16,192 weights directly. No gradients, no search, no data.
100 held-out games: 864/864 every time, zero balls lost.
https://t.co/B71LVjG3gb
4/ Evaluation: weights frozen by hash, then 100 held-out seeds.
Same architecture, parameter count and inference code as PufferAI's published trained checkpoint, which cleared 92/100 and dropped 3.2 balls a game.
With 25% sticky actions: 846 vs 297.