Labs · Field notes
Measured findings from running serious AI on ordinary hardware. Real workloads, planted-answer tests, raw numbers.
Sweep a three-point raking light across our studio replica meshes (Bearded Ruler Bust, Young Ballet Dancer, Dancing Bronze Figure) and inspect surface relief with an 8x macro loupe. Automatically falls back to a light-pooled still on Save-Data or Lite devices.
Interactive 3D exploded-view R&D CAD study of a 5-layer museum touch-kiosk architecture: optical front glass, commercial IPS panel, CNC steel frame, offline compute sled, and thermal chimney, with an orthographic 2D SVG blueprint fallback.
Field notes from K R Nariman 3D and Handloom Haat v53: SVG feTurbulence + feDisplacementMap liquid glass refraction, three-point bronze lighting, and a mobile home page cut by 63.3% with all 51 boxed card borders gone.
Measured latency, token efficiency, pass rates, and reasoning-channel traps across five Chinese frontier models and two Western baselines on three live agency engineering tasks.
Measured latency, token savings, and false-hit calibration from deploying a local ONNX MiniLM-L6-v2 semantic cache (1.84 ms median embed, 0 false hits at 0.82) and Karpathy surgical-edit rules.
Serving a 122B Mixture-of-Experts model locally on AMD Strix Halo: routing each token to 10B active parameters delivers 27.40 tok/s, running 3.46x faster than a dense 32B model.
At Q4_K_M, a dense 32B model and its full 32,768-token KV cache occupy 28.0 GB on unified memory, leaving 88 GB of headroom free while passing 5 of 5 planted verification tasks.
When you start something in the background, keep the number of the thing itself, and prove it is gone before you say it is. How three unstopped servers held the browser lock for 18 minutes, and the proof step that fixed it.
Before you approve a new version, make sure it was built on top of what is live now, not on an older copy; a version that looks right but was built on an older copy can quietly undo fixes.
What the winners share is not a look but a stance: they set the conditions of the visit, show real names and real data, and make motion answer the visitor.
A 14-hour unattended stress test of a local 120B: 5,095 self-verifying tasks, zero accuracy drift, and the empty-answer base rate measured at 1.3%.
A three-way shootout, the empty-answer reasoning-channel trap, honesty under quantisation, and two boot-time pathologies.