The classic turn-based artillery game Scorched Earth—beloved of anyone who ever shared a DOS machine in the early ’90s—has been rebuilt from the ground up as a modern Python/Pygame project with a twist: its AI opponents learn through reinforcement learning.
Dubbed MOAG (Mother Of All Games), the project shipped four commits in the first week of January, delivering a fully playable hot-seat multiplayer experience for 2–4 players with destructible terrain, a dynamic wind system, and multiple weapon types.
But the real story is under the hood. Two RL training algorithms—PPO (Proximal Policy Optimization) and DQN (Deep Q-Network)—compete to produce the best AI artillery officer. TensorBoard integration lets you watch the agents improve in real time as they learn to compensate for wind, elevation, and terrain destruction.
The heuristic AI provides a solid baseline opponent, but the RL-trained models are where it gets interesting—agents that start by lobbing shells into the void and gradually learn the physics of parabolic trajectories, one exploded pixel at a time.
“Two AI spaceships compete around a central star with gravitational pull—faithful to the 1962 original.”
— SpaceWar_AI README
In the same week, a faithful recreation of the legendary 1962 game Spacewar! landed with modern RL agents. Two AI-controlled spaceships orbit a central star under Newtonian physics with inverse-square gravity and screen wrap-around. DQN and PPO agents learn to fight, and optional LLM integration from six providers offers training-time strategic guidance. The original game ran on a PDP-1; this one runs on PyTorch and a GPU.
What do you get when you feed 41 years of hacker culture into a retrieval-augmented generation pipeline? The hacker_rag project answers that question by ingesting the complete archives of 2600: The Hacker Quarterly (1984–2024), Phrack Magazine, Cult of the Dead Cow texts, and DEF CON talk archives.
The system uses ChromaDB for vector embeddings and semantic search, Tesseract OCR for scanned PDFs, and Whisper for audio transcription of conference talks. Multiple LLM providers—Claude, OpenAI, xAI, and Gemini—or local GGUF models can then generate long-form articles about hacker community evolution, ethics, personalities, and cultural trends.
January saw documentation improvements land, refining the project’s architecture for contributors who want to help preserve and analyze this irreplaceable archive of digital counterculture.
A new AI-powered music generation tool arrived in January, combining Meta’s MusicGen neural audio model with a Strudel/TidalCycles-inspired pattern system for live-coded music composition. The tool offers two generation modes: MusicGen for higher-quality but slower output, and pattern-based synthesis for fast, real-time results.
A library-based discovery feature analyzes reference tracks to extract rhythmic patterns, while procedural audio effects—vinyl crackle, tape saturation, reverb, and lo-fi filters—add analog warmth to digital output. Breakbeat extraction rounds out the toolkit for anyone who wants their AI to produce something with genuine groove.
How many ways can you visualize a prime number? At least 33, according to Prime Plot AI, which arrived in late January with an ambitious goal: use machine learning to discover novel patterns in prime number distributions through GPU-accelerated visualization.
The project implements Ulam spirals, Sacks spirals, Klauber triangles, Vogel spirals, Fibonacci spirals, and modular arithmetic plots, among others. An evolutionary parameter discovery system optimizes visualization settings to highlight patterns that might escape human intuition, while U-Net neural networks handle prime detection on the generated grids.
Four commits in January laid the foundation, with a critical bug fix arriving in early February that corrected a detection threshold preventing the system from identifying most primes in normalized grids—a subtle issue where single primes received values of 0.33 against a threshold of 0.5.
A 2D self-driving car simulation that learns to navigate procedurally generated racetracks using three competing RL algorithms: PPO, DQN, and GRPO (Group Relative Policy Optimization). Built with Pygame and the Gymnasium framework, the simulation features realistic car physics, lidar-like 9-ray sensor arrays, and O(1) collision detection. Tracks are generated via Catmull-Rom splines, ensuring each training run faces novel geometry. Documentation landed on New Year’s Day.
The multi-AI comparison tool received a January enhancement inspired by research from Leviathan et al. (2025): repetition prompting strategies that improve response quality by structuring how prompts are repeated across evaluation rounds. The Bradley-Terry ranking system continued to evolve as a core feature for statistically rigorous provider comparison.
··· “Frustrating adversaries since the dial-up era” · GitHub: rondilley · 42 Repositories and Counting ···