Stacks of these seem sufficient to predict and generate large parts of computational reality.

CSM aiで久しぶり遊んでみたら普通に良さそうなの出てきた！！！ UEとかで飛ばせそう！！！

The move away from Lean and analytical theorem provers in AI math models is indicative of a general trend - gradually distill the analytical engine into network weights. By analogy, similar things will happen to game engines - first AI world models will augment rendering engines

In the 5-10 year horizon, an explicit 3D representation will be needed for: - Vibe coding with game engines - CAD / 3D printing (no alt) - Data engine for action-video models For 3D printing, the only other alternative is to have generalizable robot policies to execute object

Prompting detailed and editable 3D scenes is already possible. The next step is to make this as easy and fast as possible.

Figma’s journey was tough, but it paid off. In the age of AI, owning the design layer can be powerful. Most apps in production are designed in Figma and they might end up owning the vibe app/code layer distribution.

「1枚の画像」から、ここまでの3Dが作れる時代です。 Common Sense Machinesを利用したデモ。まずは、画像1枚から高精度な3Dシーンを生成。さらに、Parts with auto assemblyでパーツ毎に書き出し、Chat to 3Dを組み合わせて調整。素晴らしい精度。 もはやモデリングは「入力」次第。

Doing AC duties for NeurIPS and a reviewer literally had some prompt accidentally prefixed to their official response 😓

A casual evening walk can sometimes generate pure magic in the real physical environment (not AI generated this time :))

A neural network that wants to learn. All things considered, implementing and training a model is still personally the most enjoyable task.

We have pushed AI mesh topology quite far and its available to everyone. This is a quick test of CSM Cube with Microsoft Copilot. It is exciting to think what we can do when good 3D topology meets LLMs, meets world models.

Text, logos, and fine details have always been tough for 3D AI - but we’ve cracked it 🚀 Our latest model delivers unprecedented textured mesh quality, and it scales nicely with compute & data. Coming soon 👀

Our new 3D AI output significantly outperforms Hunyuan3D-2.5

CSMがテクスチャの品質かなり上げてる。Tripo、Rodin、Hunyuan、Hitemは大体掴めたから、次はCSMとMeshy行くか。

Here are a few more examples from our upcoming Image-to-3D model, showcasing unprecedented texture quality.

A few more texture comparisons from our upcoming model (Rodin, Hunyuan2.5).

Gemini 3 pro feels like a big step forward in terms of reasoning over complex math and ML concepts. May have helped me systematically think through a new idea - will try to experimentally verify it before sharing more.