Abstract Data

Just data stuff all day

[Utah teapot 🫖 @SkyeSharkie]

I'm making real progress with Opus 5, Fable and Sol on a very high value tool that I'm going to opensource. A @threejs webgpu based terrain generator with an API for agents to use as well. It's called Ekigar, from the sumerian E-ki-gar meaning "Temple that places the earth". I will be releasing it on my git once it gets to a good enough first state! Here's a preview of how it's going. So far, it's EXTREMELY fast, much better performance than industry standards for terrain generation.

[DynamicWang @dynamicwangs]

awplanet 3D导演台正式开源了！
基于 React、Three.js、Vite 与 Electron 构建的 AI 原生 3D 导演工作台与场景创作工具。它将可视化场景编辑、角色控制、骨骼姿势、电影化运镜、 分镜时间轴、地形雕刻、物件布置和手机虚拟拍摄整合在同一套工作流中。

项目提供第三人称游戏视角、FPV 无人机运镜和手机虚拟拍摄三种导演模式，并支持 macOS Apple Silicon 与 Windows x64 桌面端。浏览器版 Phone Pilot 可以通过手机 方向传感器和虚拟摇杆控制电脑端镜头。

awplanet is a source-available, AI-native 3D direction and scene creation studio built with React, Three.js, Vite, and Electron. It combines a visual scene editor, playable character controls, cinematic camera tools, a storyboard timeline, terrain sculpting, object placement, and a browser-based phone pilot.

项目地址：https://t.co/exDfh08rsP

欢迎大家体验及反馈！

[Argona @Argona0x]

whoever leaked this has bigger balls than sense

two researchers replaced $7,500 of human grading with $77.81 of model calls, and the number that fell out should stop anyone shipping an agent this week: the judge disagrees with itself 13.6% of the time

it also prefers whichever answer it saw first, 72% of the time. cross-judge agreement is a kappa of 0.51: the statistical polite word for guessing

this is Eval Engineering, the layer that arrives after loops and graphs, and it installs into the agent you already pay for:

- run every comparison both ways and average the two, because one production judge picked whichever answer sat in the first slot 72% of the time
- stop shipping on a single verdict: recovering the reference answer took 11 repeated trials before a majority vote landed it, and 15 on the questions that actually mattered
- report the chance-corrected agreement instead of raw match, since the gap between those two ran 33 to 41 points on the same benchmark
- never let a model grade its own family, because self-preference tracks capability upward rather than falling away as models improve
- test stability and bias together, since two judges already running in production held above 0.95 on repeat while carrying severe positional preference
- freeze the rubric wording and version it, because semantically identical templates flipped the majority outcome in a quarter of tested cases
- write the rubric as one observable pass condition rather than a bundle of scores, so a coin-flip preference cannot hide inside an average
- let plain code take every objective call, did the test pass, does the file exist, did the state change, and leave the judge only what genuinely needs reading
- make the verdict do something structural to the run in flight, reject the handoff, swap the node, quarantine the branch, or you have built a dashboard

the catch is that your judge is a measurement instrument nobody calibrates: single-trial judging is too noisy for anything you would stake money on, and the same judges reordered by as much as 14 positions depending on which benchmark you asked

so the loop has to converge on the spec rather than on the score, because optimise against a judge long enough and the agent learns to look right instead of being right

bookmark this, the full build with the five evals to start on and the gate that lets the fleet merge its own work is written out in the article ↓

[Hardy @Yun_HDY]

when ardy dropped i said this is what an ai native engine looks like

so i wired it into blender

walk down the stairs, cross the gap, land
that was the whole input. zero keyframes

[ITBald @hamster_wat]

A 3D @threejs building generated with img2threejs and Opus 5. The new Opus isn’t magic, but it handles details a little better than GPT-5.6 Sol.

#threejs #webgpu

[Chris @ChrisGPT]

Claude Opus 5,

I think we’ve reached the inflection point where AI can produce indie quality assets for video games.

This is after my second prompt, just simply telling it to increase the quality.

Some of these angles on the car are truly stunning and I can 100% say they look better than a lot of Indie games I’ve recently played.

[爱丽丝呀！ @BTCqzy1]

为什么你的 Codex 5.6 搓出来的 3D 像一堆方块，别人的是产品级？

之前给大家分享的一个宝藏开源项目img2threejs，接入 Codex 5.6 后，就能完成：

2D 图像 → 结构分析 → Three.js 程序化模型 → 浏览器实时交互

它真正拉开差距的地方在于，给 AI 加了一套完整的���序化建模操作系统：

1️⃣AI 深度拆解轮廓、比例、细节特征

2️⃣从 Blockout 到灯光、交互，步步��进

3️⃣浏览器渲染对比，不完美就修改

4️⃣保留运行时层级、轴点和插槽

所以真正拉开你和产品级差距的，是你有没有把建模标准、检查点和返工机制，像 img2threejs 一样，写进你的提示词。

这才是 Codex 5.6 真正开挂的玩法！

我把我的promote放在评论区，可以参考～

项目地址：https://t.co/fKNYoeLe3A

[Abstract Data @abstract_datum]

@LunarResearcher @grok youtube link

[Rina @irinatoxi]

The most overpaid part of product CGI is about to become dangerously automatable.

A shot like this looks “simple” until you remember what it actually takes: modeling the form, breaking it into parts, animating the assembly, tuning the glass material, getting the reflections right, cleaning the motion, and making the final frame feel premium instead of cheap.

That is why videos like this are so interesting. They are not just pretty renders. They are a preview of the kind of structured 3D work AI will eventually eat first, because the whole thing follows a logic: parts, motion, materials, lighting, composition.

The scary part is not AI making one lucky image of a bottle. The scary part is when it can reliably build the full editable scene, animate every piece into place, inspect what looks off, and keep iterating until the ad actually feels brand-ready.

When that moment arrives, a huge chunk of “high-end 3D product work” stops being manual craftsmanship and starts becoming art direction. And honestly, I don’t think we’re that far from it.

[Hugging Apps @HuggingApps]

ABot World 0.5B is out! we have now world models at home, running in real-time on consumer GPU

feed it an initial image, steer it with with your keyboard 🌎

you can run it locally or play now on spaces
▶️ https://t.co/0JUL2rRspW

[FHILY👑 @Oluwaphilemon1]

Claude Opus 5 is honestly insane compared to Kimi K3, GPT-5.6, and Claude Fable 5

Claude Opus 5 built a playable Portal-style game in Unity in less than a day.

✅ Working Portal Gun
🌀 Real portal physics
⚡ Lasers and gels
🎮 Core gameplay mechanics fully playable

The biggest unlock wasn’t just the model.

By connecting Claude directly to Unity through a free MCP plugin, it could build, test, iterate, and refine gameplay systems inside the engine, while replacing placeholder assets with AI-generated 3D models along the way.

We’re rapidly approaching a world where describing a game is becoming the hardest part of making one.

[Chris @ChrisGPT]

I asked Claude 5 Opus to generate me the best game graphics it could!

I wanted to create a car on a dirt trail demo and see how good it was at crafting car graphics without any textures.

Everything you see here is 100% crafted from the model itself.

[Johnathan Chiu @johnathanchewy]

1/ Appreciate everyone who showed interest and reached out! The plan is to release code and weights in the next few weeks after additional improvements.

I wanted to share more about my method here in the meantime and happy to discuss with anyone who wants to chat!

⬇️ See below

[darkzodchi @zodchiii]

Anthropic engineer:

"You don't need better prompts. You need graph engineering: memory that stays, so your agent never forgets anything."

In 28 minutes he shows what Anthropic does differently, how to build and structure work with agents.

This beats any paid agent course I've seen.

Watch it, then read the graph engineering guide below 👇

[Tore Lervik @toreler]

Blending the edges between different environment meshes in UE5. Removes so much of the noise! 🤩

PS: The @fab Summer Sale ends tomorrow at 11:59pm ET!
https://t.co/asG19GK43K

#gamedev #unrealengine

[逛逛GitHub @guangGitHub]

终于有一个可以复刻 Apple 设计风格的 Skill 了。
Apple 的一些视觉风格，给人的感觉就是很舒服。

apple-design ，它的来源主要来自 Apple 的 WWDC 设计演讲，尤其是流体界面、音频触觉反馈、字体细节、伟大设计原则这些内容。

作者把这些经验整理成 17 条设计和动效原则，再转换成 Web 里能落地的写法。

Skill 最核心的理念有那么几个。

比如界面要从用户当前的动作里长出来。

它不追求把动画做得花，但是要让界面始终保持可控。

它还包括材质和层级。

导航栏、工具栏、弹层这些浮动区域，可以用 backdrop-filter 做出半透明的层，但不能一层一层叠到看不清。

使用方式很简单。

npx skills@latest add emilkowalski/skills

装好之后，就可以在做界面时调用 apple-design。

[George from 🕹prodmgmt.world @nurijanian]

This legit feels like I have a McKinsey consultant team in my pocket every time I start a new problem space, area, or investigation.

The reason is simple: I’ve scoured the literature—from Minto's Pyramid Principle to other consultant-written books—and combined that with what I’ve learned from doing it myself and watching peers do it, too.

I’m hopeful people will find it useful.

[Jake Fitzgerald @earthtojake]

text-to-cad just crossed 10,000 stars

you can use it to generate:
- STEP files
- URDF, SDF and other sim artifacts
- 3D mesh files like STL, 3MF and GLB
- gcode for 3D printing
- DFM checks for popular services like sendcutsend

100% free and open source ✌️

[Matt Shumer @mattshumer_]

3 days ago, I posted a game Claude built from a single prompt.

3.8M views. It looked so real, folks were convinced it was fake.

It's real, driven by a technique I'm calling the Gauntlet Loop.

Here's how to run one. It works for far more than games:

https://t.co/EGOLEUxQtQ

[3DxDEV @3DxDEV7]

Quick Menu by passivestar. This is a free, all-in-one, node-powered add-on that bundles a whole pile of workflow tools into one slick menu.

👇Link In Comments