Thinking
Essays
Long-form notes on the systems we build and the principles behind them.
Vision-Encoder Behavioral Fingerprints of Image-to-Image Generative Models: A Training-Paradigm-Driven Taxonomy of Six Commercial APIs
A reference-anchored forensic study of six production image-to-image APIs (gpt-image-1, Gemini 2.5 Flash Image, Flux Kontext, SDXL img2img, SD3 img2img, Qwen Image Edit) under content-adaptive sub-JND perturbation. Across a 3,588-call corpus, the systems partition into two image-invariant behavioral bands whose discriminating variable is training paradigm rather than architecture family, enabling 51.4% six-way attribution from a single output plus a clean reference and exposing differential perturbation survival across deployed commercial systems.
The Hostility of Abundance
Why we're manufacturing scarcity back into the world — and why dating apps, dumb phones, traditional religion, and the disappearance of moral stories are the same trend.
AI Training Data Has a Provenance Problem
The gap between how AI companies source training data and what creators actually consented to — and how Stelais creates the only way out.