Essays on AI agents, swarm intelligence & systems engineering.
When AI Becomes Its Own Scientist
Inside the Evolution Arena and the rise of autoresearch.
An AI agent that proposes its own experiments, runs them against a live 2D survival simulation, scores the result with a hard mechanical metric, and commits or reverts on its own. No vibes, no subjective review — just a ratchet that only moves forward.
When Swarms Write Code
How particle swarm optimization escapes the local-minima trap in ARC-AGI.
Standard LLM agents get stuck in local minima on ARC-AGI. The fix: swap the single-agent loop for a PSO-governed swarm of specialized LLM particles with a continuous fitness function that rewards near misses. The swarm provides strategy; the LLM provides syntax.
Stop Wrestling with Boilerplate
Local Tinker — a clean API for local LLM fine-tuning.
A Tinker-style API for LoRA fine-tuning of 1B–13B LLMs on your own GPU. Four primitives — ServiceClient, TrainingClient, forward_backward, optim_step — cover SFT, DPO, PPO, and GRPO without the usual HuggingFace + PEFT + bitsandbytes boilerplate.
Building an AI That Masters Snake
A deep reinforcement learning project from scratch.
How I built a Snake AI with Deep Q-Learning in PyTorch. A neural network, a shaped reward signal, and a lot of virtual trial and error — no hand-coded strategy, no search algorithms. Averages 44 points over 200 games, peaks at 75.
Two AIs, One Loop
Building a self-improving code agent.
A two-agent architecture — one Claude planning, another implementing, with git diffs and test results flowing between them — captures most of the value of multi-agent coding systems while avoiding their complexity. What I built, and what 50+ research papers say about why it works.