10.11.26 | 2D Vehicles, The Lightbulb Computer, City building game resists development

SPEAKER_1 (0:00)

Welcome to the Hacker News Highlights, where we explore the top 10 posts on Hacker News every day. Today, we dive into recreating GTA-inspired 2D vehicle physics, projecting ambient computing through lightbulb-shaped devices, and simulating San Francisco housing resistance. Let's dive right in.

Title, 2D Vehicles. Source, Pat Kerr. The project recreated a 1996 physics demo that later became the vehicle system for Grand Theft Auto. It ran a 2D rigid-body simulation and added simple car behavior by applying forces and resistance at the tires. The JavaScript version offered car, spaceship, and brick modes, along with gravity, barriers, and camera controls. The project explained how the physics worked and showed how simple approximations could create convincing handling. In the comments, the sentiment was largely positive, with users praising the demo's playability, nostalgia, and connection to early Grand Theft Auto games. Users debated whether using artificial intelligence to help create diagrams weakened the project's hand-coded history. Supporters said the author had written the core text and recreated the original logic, while critics found the generated diagrams unclear and overly generic. Overall, users saw the project as a fun and effective demonstration of simple game physics.

Title, The Lightbulb Computer. Source, Guillaume Ardoux. The project described a speculative computer that combined a projector and computer vision inside a lightbulb-shaped device. It projected recipes, maps, photos, schedules, and controls onto household surfaces, using voice commands and pointing. The article argued that ambient projection could offer more social and practical interactions than phones, tablets, or headsets. The prototype used a normal projector, a webcam, and custom software, and the author said privacy would require local processing and physical camera controls. In the comments, the sentiment was largely enthusiastic, with users praising the idea as a more social alternative to glasses and screens. Users debated whether projectors could provide readable images in daylight, and whether the proposed hardware could fit inside a practical light bulb. Supporters pointed to cooking, shared maps, and hands-free interactions as strong uses, while critics raised concerns about shadows, heat, latency, camera coverage, and privacy. Overall, users saw the concept as exciting, but still far from a dependable consumer product.

Tidal, a city-building game in which the city would prefer you didn't. Source OverDot Pizza. The project, Discretionary Review, was a San Francisco housing simulator about trying to build homes through permits, hearings, appeals, lawsuits, and neighborhood opposition. It drew its rules and case counts from public records, then simplified local politics and housing economics for play. The game used bureaucratic complexity as both its main mechanic and its joke. In the comments, sentiment was divided between praise for the simulation and frustration with its difficult interface. Users debated whether the complexity effectively showed the obstacles facing housing developers or made the game too demanding for a short educational project. Supporters said the layered systems made the city's competing interests feel real, and players shared strategies for building thousands of homes. Overall, the game succeeded as a memorable simulation, but its audience and purpose remained unclear.

Title, Build Your Own Decision Model, Source, Another Dev's Two Cents. The post explained how to turn a language model into a fast decision model by restricting its output to a fixed set of choices. It used Quen version 3.1.7 billion to classify multiple choice questions, reaching about 59% accuracy on Common Sense QA without fine-tuning, and about 62% after fine-tuning. The post also showed that raw token probabilities were badly overconfident. Temperature scaling improved calibration by making confidence scores better match actual accuracy. A GitHub project provided scripts for building, evaluating, fine-tuning and calibrating similar models. In the comments, sentiment was divided between excitement about vast, cheap, zero-shot classification and skepticism about the new terminology. Users debated whether decision models offered a real advance over traditional classifiers and ordinary language models. Supporters cited simpler APIs, lower costs, faster responses and useful applications such as text filtering, game logic, and reducing token use in larger workflows. Overall, users saw value in the approach but questioned its novelty, calibration claims, and long-term advantage.

Title, Talorys, a self-hosted personal AI agent on Cloudflare's free tier, Source, GitHub. The project was an open-source personal AI assistant that ran inside a user's Cloudflare account. It offered chat, persistent memory, tasks, notes, projects, reminders, and scheduled automations through Cloudflare pages, workers, durable objects, and workers' AI. A command-line installer handled deployment while the project stored data in a SQLite-backed durable object and included password protection, backups, updates, and no developer-run telemetry. The free tier had usage limits, and Cloudflare still processed AI requests and controlled the underlying platform. In the comments, sentiment was divided between appreciation for the implementation and criticism of the self-hosted label. Users debated whether running software on a VPS or Cloud platform counted as self-hosting, or whether it required personally-owned hardware. Supporters praised the simple deployment, useful durable objects architecture, and the possibility of replacing Cloudflare's AI service with a local model. Critics warned about vendor lock-in, changing free-tier limits, slow workers' AI responses, privacy, and unexpected Cloudflare bills. Overall, the project was seen as useful, but self-managed or Cloud-hosted was judged more accurate.