Notes from the trenches on AI engineering, LLM apps, and the full-stack work that holds it all together.
The SelfRisingRobot project trains a self-righting policy in MuJoCo with PPO, exports it to a C header, and runs inference on an M5Atom with two servos. Codex helped scaffold the sim loop. Sim-to-real without cloud GPUs at runtime.
Cloudflare Workers AI gives you 100,000 free image generation calls a day. I built a simple worker that turns text prompts into images with Stable Diffusion, no GPU bill.
A hand-picked directory of free AI courses, math resources, datasets, and tools, annotated so you know what's worth your time and what to skip.
Goodfire's Silico platform can forecast which behaviors a preference dataset will teach a model, with R² around 0.9 in their studies. Filter clusters before DPO instead of reverse-engineering failures after training.
Karpathy's moving to Anthropic to lead a team using Claude to build better Claude. It might be the most interesting hiring move of 2026.
A new arXiv study maps correlated name priors in Claude, GPT, and Gemini outputs. Fictional personas like Elena Vasquez and Marcus Chen appear across AI-generated sites and 1,655 backdated Zenodo records with real DOIs.
26 lessons, 52 quizzes, Scikit-learn projects, and Jupyter in VS Code. Microsoft's MIT-licensed curriculum is the on-ramp I still send before deep learning rabbit holes.
SkillSpector joins a crowded field of agent skill scanners with 70+ vulnerability patterns and optional LLM analysis. New research shows packed and obfuscated skills still bypass most static tools. Scan first, sandbox second.