Urgent.News

What's breaking now, across thousands of outlets.

AI

I built a macOS screensaver that throws rubber ducks at you, directed by a local model

Mac Attack watches the room with the Mac camera, turns each person into a cartoon character, and fires harmless toy effects at them: bubbles, rubber ducks, tomatoes, confetti. It runs as a normal app and as a real macOS screensaver. Everything stays on the Mac. Apple's Vision framework produces body rectangles only (no face recognition), frames are processed and released, and nothing is uploaded.…

A developer created a macOS screensaver called Mac Attack, which throws harmless effects like rubber ducks at people in the user's room. The screensaver runs as a normal app and as a real screensaver, with all processes happening on the Mac. Apple's Vision framework is used to detect and turn people into cartoon characters. The developer, Laya, is a non-autoregressive typed-decision model that processes body rectangles and provides calibrated probabilities for various actions. By using sampling instead of the top answer, the game feels more dynamic and unpredictable.

Three key measurements were taken: 1) sampling beats argmax, 2) policy belongs in code, not in the prompt, and 3) measure the model you're actually running. Sampling from the calibrated distribution made the game feel more alive, while rewriting the questions to make rare options less likely helped prevent the game from becoming stagnant. Additionally, splitting the model's input into smaller questions increased the diversity of responses.

The challenge of creating a real screensaver was overcome by using a launchd agent to own the camera and serve anonymous person boxes on 127.0.0.1. The screensaver then polls the agent at 10 Hz, runs the game engine, and renders the effects. This architecture ensures the camera is only used when necessary and the screensaver remains fully functional even when the main application isn't running.

Mac Attack is currently an unsigned build, so Gatekeeper will prompt users to open it. The entire source code, architecture notes, and a detailed account of what worked and didn't during development can be found on GitHub.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at dev.to →

More in AI

FinTrack-AI : From Excel Sheets to AI Financial Companion

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend What I Built What I Built I built FinTrack AI for my father and uncle.

  • FinTrack AI is AI-powered finance tracker created by father and uncle
  • Users upload Excel/CSV data, no manual entry required
  • AI inference runs locally via Ollama, preserving privacy and control

InterviewBuddy AI

🤝 InterviewBuddy AI --- A Local AI Interview Partner for a Friend This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend What I Built I built InterviewBuddy AI , a…

  • InterviewBuddy AI is AI-powered mock interview partner for data analyst roles.
  • Users upload resumes, choose job roles, and select difficulty levels.
  • AI provides personalized interview questions, feedback, and improvement plans.

We re-tested 69 skills on harder inputs. 16 failed. Here is what broke.

Disclosure: we make agent skills (see the end). The checklist below is ours and needs no purchase. Written with AI assistance. Every skill we ship had already passed one run on a real public input.

  • 69 skills re-tested with harder inputs
  • 16 skills failed harder input test
  • Common defects: unsure input as fact, contradictions

More from Sunday 4 October →