How I solved the background removal problem for AI-generated animated WebP tiles, and why color detection and chroma key both fail.

Christian, late-diagnosed autistic, technologist.

How I solved the background removal problem for AI-generated animated WebP tiles, and why color detection and chroma key both fail.

LLMs are autocomplete engines trained on human culture. Sci-fi authors spent decades writing how robots fail. That's the training data. That's the completion. This isn't a bug, it's the design working exactly as intended.

Four years of AI image generation, three platforms, one stubborn cartoon bee, and the declarative pipeline I built to stop the drift.

The AI debate is forcing us to define intelligence. Every definition proposed, when applied consistently, ends up drawing a line through humanity itself.

Hallucination isn't a bug in LLMs — it's the mechanism. The real problem is misallocating non-determinism, and why "SKILLS.md is enough" is exactly backwards.

AI didn't create the problem of empty content. It made it cheaper. 3,000 years ago Qoheleth had a word for it.

Most people misunderstand what generative AI actually does. It doesn't add quality — it multiplies what's already there. Which means the people it serves best are the ones who least needed it to begin with.

Using Claude & MCP servers to enhance genealogy research.

MyHeritage in Color allows to fine-tune automatically colorized and enhanced photographs