Shibavoxel

Geopolitics game theorist. Voxel Art, Tokenisation, Chainlink RWAs. Modeling G3 world as H200 hits China. Builder. Contrarian.

[Shibavoxel @VoxelKBKR]

@cixliv @nicdunz

[Shibavoxel @VoxelKBKR]

@ash_twtz Trust me bro benchmark 🤣

[Shibavoxel @VoxelKBKR]

@louszbd Did GLM-5.3-Flash generate production ready topology and baked maps for the vehicles, or is the result still a high poly visualization mesh that would need a human artist to retopo and bake before it could leave the look dev stage?

[Shibavoxel @VoxelKBKR]

@ImtiazMadmood 🤣🤣🤣

[Shibavoxel @VoxelKBKR]

@swyxy4484 🤣

[Shibavoxel @VoxelKBKR]

@NFT_Chen

[Shibavoxel @VoxelKBKR]

@xikhar @fawnunwrap @rosspeili

[Shibavoxel @VoxelKBKR]

@hajime2ai

[Shibavoxel @VoxelKBKR]

@youling6379

[Shibavoxel @VoxelKBKR]

@OpenAIDevs

[Shibavoxel @VoxelKBKR]

@JeSuisYasu

[Shibavoxel @VoxelKBKR]

Anime desktop pets just landed!
SD Gundam + Chibi Asuka — idle, wave, jump, get sad, stare in 16 directions… the whole vibe.
Codex-compatible v2 sprites, ready to install.
https://t.co/liaxtQ45O9

[Shibavoxel @VoxelKBKR]

@elonmusk 🤣🤣

[Shibavoxel @VoxelKBKR]

This is a really cool project.
Right now AI world gen is basically:
-awesome for greyboxing / concept blockouts
-great for skyboxes + distant scenery
-insane for VR splats that feel real

But the moment you need doors that open, weapons that collide, or lighting that actually behaves… you’re back in “hello Blender my old friend” territory… artists still have to craft, retopo, and bake real meshes.

So yeah: AI gets you 80% of the vibe fast… the last 20% is still very much human craftsmanship.

[Shibavoxel @VoxelKBKR]

@Dr_Singularity GPT-5.5 still suggested I walk to the car wash because it’s only 2 minutes away, instead of driving the car that actually needs washing. 😅

[Shibavoxel @VoxelKBKR]

@r0ck3t23 @Global_Impulse1 AI still obeys the law of supply & demand. Whether it’s a US or Chinese model, they all need DCs, power, cooling, networking, and compute. “Cheap AI” is often subsidized by public or private capital the infra bill doesn’t magically disappear.

[Shibavoxel @VoxelKBKR]

Running a 235B model and running a 235B model well are two very different things. The real question isn’t whether it fits into memory. It’s tokens/sec, latency, context length, agent workflows, and concurrent workloads. We’ve seen this movie before with every personal AI supercomputer launch. Capacity is not the same as performance.

[Shibavoxel @VoxelKBKR]

📡 This week in homelab AI infra:

New research digest available

#homelab #llminfra #SRE #buildInPublic

[Shibavoxel @VoxelKBKR]

@elonmusk Inner Chongqing might be the ultimate stress test for autonomous driving, and Tesla FSD seems to be handling it surprisingly well.

[Shibavoxel @VoxelKBKR]

📡 This week in homelab AI infra:

New research digest available

#homelab #llminfra #SRE #buildInPublic