Articles

How I stopped fighting with sequential model loading on a multi-GPU VM and built a simple job queue with uvicorn workers instead.

I built a system that streams Minecraft from L4 GPUs in Kubernetes pods, with an AI agent inside the container that writes Java plugins, compiles them, and tests them in-game.

On using a separate vision model (Gemini) for visual QA instead of letting the same agent judge its own 3D models. Fresh eyes matter, even for LLMs.

Comparing Tesseract, AWS Textract, and Azure Document Intelligence for scientific paper OCR. Plus notes on semantic chunking and image retrieval.

Neural networks don't tell you when they're guessing. We tested a cheap way to fix that by looking at the model's internal state, and it outperformed the expensive standard methods.

What if a model only explored alternatives when it was unsure? We tested this against standard approaches and found it recovers errors with less compute.

Chain-of-thought reasoning gets perfect scores during training, then collapses to 0% one step outside its comfort zone. We found an alternative that degrades gracefully instead.

AI agents confidently walk off cliffs they can't see. We ran three sets of experiments showing that self-awareness, not just raw capability, is what's missing.

What if a neural network could replace a game engine? I recorded thousands of frames from a 3D robot fighting game and trained a model to generate the next frame from controller inputs.