Call me Cassandra
Three years holding this back. Done holding.
All Posts
18 articles
Turning a face into a chibi: a tour of LAM, Gaussian Splatting, and what happens when you push on the mesh underneath
A few weeks ago the only thing we wanted from a "talking-head avatar" was for it to look like the person in the input photo and to blink at roughly the right moments. By this week we wanted to grab th
On PMs and LLMs: Forwarding the Work of Thinking
He had been asked a question and he had not bothered to answer. Between a wife who could ask and a fish who would grant, he had only to think — and he had not learned to.
Of Solitary Vice
Most work that feels productive is self-referential, and the small warm room is unlikely to tell you so. An essay on the alchemist's tower, the journals at the foot of the hill, and the visitor whose advice is absorbed, not refused.
This Is the Lucky One
Most decisions you celebrate were wrong, and nothing in your week is going to tell you which ones. An essay on feedback loops, falsification, and the discipline of being made to find out.
From a single crash to a public lint plugin: encoding RN/Expo tribal knowledge as oxlint rules
A production Android RN/Expo app crashed on launch with the canonical message:
Four probes, three falsifications, three fixes — accelerating PersonaLive
PersonaLive is a recent portrait-animation diffusion stack — reference UNet, denoising UNet with a temporal module, motion encoder, pose guider, VAE — that ships at ~10 FPS on an RTX 5090 with stock S
From Essay to Operating System: A Framework for Working With AI Without Losing What You Need to Know
A day after publishing Renting Competence, I sat back down with the same question and pushed on it harder. The essay had a thesis: if you let an LLM hold the model of your system, you keep shipping bu
Renting Competence: A Framework for Working With AI Without Losing What You Need to Know
Here is a question I have been asking developers a lot lately:
Why the squint slider needed an ArcFace anchor — the v1c → v1k failure ladder
From FluxSpace edits to a working glasses slider — what five days taught us
Composing a smile without losing the face it belongs to
Looking for AU atoms in Flux's attention cache
FluxSpace, End to End: One Recipe, Two Axes, Six Demographics, and a Geometry That Travels
Date: 2026-04-21 Series: Follows Demographic-PC Extraction, End to End. That post ended with a warning: our ridge-direction edits collapsed into solid colour or noise well before the published FluxSpa
Demographic-PC Extraction, End to End: Ridge vs Prompt-Pair, Two Kinds of Cliff
Date: 2026-04-20 Series: Follows Perception Before Training. That post framed a six-level perception curriculum gated on a Level-0 engineering task: build a demographic subspace in Flux conditioning s
Perception Before Training — A Curriculum Interlude Between Parts 2 and 3
Date: 2026-04-20 Series: Sits between Part 2: The Math of Pattern Preservation and whatever Part 3 turns out to be. Part 2 argued for a training pipeline that preserves patterns from text through face
Faces as a Data-Mining Puzzle — Part 2: The Math of Pattern Preservation
Date: 2026-04-16 Series: Part 2 of 2. Part 1 here covered positioning — why faces, why a game, why co-discovery. This post covers the math: how we know when a built pipeline actually preserves the pat
Faces as a Data-Mining Puzzle — Part 1: The Case for an Unlikely Interface
Date: 2026-04-16 Series: Part 1 of 2 — positioning. Part 2 covers the math of pattern preservation. Audience: technical readers who know embeddings, neural networks as black boxes, and basic ML, but n