Qwen3.6 35B Q4 ran nearly 3× faster than Q6 with only a 0–2% quality tradeoff in our tests.

1× RTX 3090 also beat 2× when the model fit. But for a 235B Q2 hybrid load, the second card made the whole workstation far more usable.

Receipts: youtu.be/W9WyRVbb7Mc

I ran 120 matched image-to-3D tests through Pixal3D and TRELLIS.2. I checked what survived: defining features, thin structures, part relationships, and finishing. The result is an inspection method, not a universal winner. youtu.be/asw3C0W68JE

A four-year-old and her little sister wrote and directed a 16-minute fantasy movie. I built the AI-assisted production workflow; they supplied the story, characters, and decisions. Dragon Snare is imaginative, ambitious, and completely theirs.

Watch: youtu.be/TXcAnCl0i9Y

Models Guide - Public Episode 04

Five tests for programmed interface compression, plus a clean way to falsify the claim: remove the shared codebook and see whether the compact message still reconstructs correctly.

A three-minute HMIC explainer:
youtu.be/Pg8bVh-PR7I

HMICsource is live. Ep.1 introduces Programmed Interface Compression: when shared state carries enough meaning that later messages can shrink toward near-nothing. The definition is one sentence. The test is falsifiable.

Watch: www.youtube.com/watch?v=rPfz...

HMICsource will share compact research notes, working artifacts, and clear boundaries around what the evidence does or does not show. Current projects include Programmed Interface Compression, code-learning games, and human/AI production workflows.

HMIC is an independent research studio exploring a simple question: how much useful structure can we compress into an interface before the interface becomes the program?
We build prototypes, test them in public, and share what survives. Starting here.