Most of what I build starts the same way: something assumes I can't get in, and I'd like to know why.

Sometimes it's an NVIDIA card. Nearly every interesting ML project has one baked in, and I have a Mac, so I take software written for hardware I don't own and make it run anyway, which is a worse hobby than it sounds. That's an int8 Metal backend for CTranslate2 (Whisper beam decode, 39s → 4.6s), an MPS lane for finetrainers, and Open-Sora 2.0 generating video on a laptop.

Sometimes the wall is a bug nobody can see. My favorite so far: in bitsandbytes (the k-bit quantization library most low-memory fine-tuning runs on), anyone training with the Lion optimizer and weight decay turned on wasn't getting the weight decay they'd asked for. No crash, no warning, just a setting that didn't do what it said, and models that came out quietly different from the algorithm on paper. It survived because the test suite only ever ran Lion with weight_decay=0, the one setting where the bug is invisible. Merged, then merged again when the same mistake turned up in a second kernel.

And sometimes the wall is the language. ai.ericeaglstun.com is where I think out loud about AI in the plainest language I can manage. Not tips and tricks, not hype, not doom, but the philosophical end of it, with a glossary underneath for when the vocabulary is what's in the way. Almost nobody fully understands how this works, and pretending otherwise helps no one. fable-mvp.tech is a séance you can run: pick a model, turn up the temperature, watch it lose signal. dot-claude is the public slice of my Claude Code setup, including a skill carrying forward the Beagle Bros sensibility: documentation from 1983 that left an eight-year-old understanding maybe half of it and wanting the other half.

Twenty years of web before this, and an art degree before that, which is my excuse for caring how a thing feels and not just whether it runs. The music, the synths, and a guitar wired to a Teensy live over at ericeaglstun.com.