Writing
Long-form on how language models represent the world — and on crypto market structure and Hong Kong virtual-asset regulation. Several of these began as Chinese essays and were rewritten, not translated.
- The Same You, Scored Differently
Changing two backend settings tripled a model's ARC-AGI-3 score without touching a single weight. Capability is a joint product of the model and the harness — and the machine now reading your résumé prefers sentences it wrote itself, 67 to 82% of the time.
- Scale Does Not Fill an Interface Blind Spot
Freeze CLIP's vision encoder, change nothing about its information, and retrain only a linear map on the text side: binding accuracy jumps from 0.58 to 0.95. The blind spot was never in the encoder. It was in the interface — and no amount of scale touches it.
- Understanding and Fitting Share the Same Gradient
Cross-entropy is the only judge in LLM training, and it cannot tell a model that understands from one that merely fits — both produce the same token and the same gradient. Which means understanding is never selected for. It only ever sneaks in disguised as compression.
- The World AI Can Compute, and the World It Can't Say
AlphaFold won a Nobel Prize without leaving behind a single readable equation. Three hard constraints suggest that for a large class of high-dimensional systems, compressing the law back into symbols a human can hold is not hard — it is unavailable.
- The Native Tongue Is Not in the Vocabulary. It's in the Manifold.
"Which human language suits AI best" is the wrong question. Human language is lossy compression built for a serial human channel the model does not have. Force a reasoning model into a single language and its maths accuracy drops 5.6 points.
- Claude Already Dreams. It Only Got the Boring Half Right.
Human sleep runs two systems: slow-wave consolidates and deletes contradictions, REM recombines and preserves them. Anthropic built the first half and called it auto-dream. The half they left out already has an engineering name — hallucination — and we spend everything we have suppressing it.
- Tokens Aren't Dying. They're Ceding the Middle.
A structural audit of discrete tokens against continuous embeddings using three independent mathematical tools — information theory, information geometry, and representation theory. The verdicts disagree, and where they disagree tells you the shape of the next architecture.
- Plato Finally Found His Wax Tablet
Anthropic cut Claude's blackmail rate to zero not by showing it rules but by feeding it millions of tokens of synthetic fiction about a psychologically healthy AI. That is not alignment engineering. It is fiction authorship — and it is the thing Plato designed 2,400 years ago and could never build.
- AI Didn't Kill Hierarchy. It Moved House.
Block says AI will replace its management layer. Prussia ran that experiment in 1808, won three wars with it, and watched it fail catastrophically in 1914 — then had the surviving staff officers write the history that exonerated the staff system. Sovereignty does not vanish. It relocates.
- Taste: A Two-Hundred-Year-Old Consolation
Silicon Valley has converged on taste as the skill AI cannot replace. Handloom weavers said it about hand-feel, Arts and Crafts said it about the maker's judgment, typesetters said it about the eye. Every time, the narrow technical claim was true. Every time, the top tier fed 1–2% of the trade.
- The Door Nobody Knew Was Open
Two uncoordinated engineering decisions accidentally gave GPT-4 a shortcut for reading meaning out of Chinese character shapes. Fixing the tokenizer closed it. The same mechanism is retiring SWIFT MT — and it is how infrastructure quietly deletes capabilities nobody knew they had.