Itch.io is confronting a sharp rise in game uploads driven by the spread of generative-AI tools. From August 3 to October 30, new game pages grew from just over 3,000 to more than 6,000 a day, while downloads and purchases did not rise comparably. The platform says the growth mainly comes from first-time uploaders, rather than mass uploads by individual accounts. It has removed bulk-uploaded near-duplicate games from search because such behavior violates its guidelines, and says it has deleted “a couple terabytes” of AI-generated low-quality material.
The response combines stronger moderation with disclosure and user choice. Itch.io now requires creators in tabletop games, books, comics, and soundtracks to disclose generative-AI use, including when AI-made material was later edited. The platform is extending AI-disclosure filters to search so users can exclude games made with these tools. Some users want a total ban, but filtering falls short of that demand; enforcing a blanket prohibition could also impose substantial moderation costs, and some developers consider AI useful for coding even if they reject AI-generated artwork. The measures aim to reduce clutter while preserving a place for creators who use AI.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
A Korean B2B company runs 17 Claude Code bots for $100 a month, using an existing office Mac and a Claude Max subscription with no API spending. Only three bots use a model; the other 14 are Python scripts, and model-powered tasks call the Claude Code CLI. (botsinproduction.com)
IKEA’s AI assistant Billie handles about 47% of customer inquiries, saving Ingka Group nearly €13 million. Ingka retrained roughly 8,500 call agents to help customers design rooms, addressing requests that require human judgment. (pymnts.com)
Effective AI prompting relies on programming skills: defining the starting point, goal, steps, edge cases, and success criteria. Engineers often get better results because they have practiced this structured problem-solving, including on noncoding tasks. (brethorsting.com)
Rotary position encodings for graphs use Laplacian eigenvectors to represent node positions and create relative attention scores without an all-pairs bias matrix. This makes the method compatible with linear attention, potentially allowing graph transformers to scale to larger graphs. (idlemachines.co.uk)
Nvidia CEO Jensen Huang says the company supports both open and closed AI models because it profits when its hardware powers either approach. He sees open models tailored with specialized knowledge as a major opportunity for businesses. (theregister.com)
AI model weights are large, fixed files that many users need in exactly the same version, making them well suited to peer-to-peer torrent distribution. Content hashes give each model version a unique address, letting a shared swarm distribute identical files across many peers. (pirateface.co)
Today's best Hacker News stories, summarized and screenshotted, one email a day.