fseraph

Game Developer

[Jackson Atkins @JacksonAtkinsX]

My brain broke when I read this paper.

A tiny 7 Million parameter model just beat DeepSeek-R1, Gemini 2.5 pro, and o3-mini at reasoning on both ARG-AGI 1 and ARC-AGI 2.

It's called Tiny Recursive Model (TRM) from Samsung.

How can a model 10,000x smaller be smarter?

Here's how it works:

1. Draft an Initial Answer: Unlike an LLM that writes word-by-word, TRM first generates a quick, complete "draft" of the solution. Think of this as its first rough guess.

2. Create a "Scratchpad": It then creates a separate space for its internal thoughts, a latent reasoning "scratchpad." This is where the real magic happens.

3. Intensely Self-Critique: The model enters an intense inner loop. It compares its draft answer to the original problem and refines its reasoning on the scratchpad over and over (6 times in a row), asking itself, "Does my logic hold up? Where are the errors?"

4. Revise the Answer: After this focused "thinking," it uses the improved logic from its scratchpad to create a brand new, much better draft of the final answer.

5. Repeat until Confident: The entire process, draft, think, revise, is repeated up to 16 times. Each cycle pushes the model closer to a correct, logically sound solution.

Why this matters:

Business Leaders: This is what algorithmic advantage looks like. While competitors are paying massive inference costs for brute-force scale, a smarter, more efficient model can deliver superior performance for a tiny fraction of the cost.

Researchers: This is a major validation for neuro-symbolic ideas. The model's ability to recursively "think" before "acting" demonstrates that architecture, not just scale, can be a primary driver of reasoning ability.

Practitioners: SOTA reasoning is no longer gated behind billion-dollar GPU clusters. This paper provides a highly efficient, parameter-light blueprint for building specialized reasoners that can run anywhere.

This isn't just scaling down; it's a completely different, more deliberate way of solving problems.

[Jakub Boksansky @boksajak]

I put all of my Ray Tracing Gems chapters in one place where you can download them, extracted from books into separate PDFs: https://t.co/AOGFePRgoG
There is:
1. "Ray Traced Shadows"
2. "Reference Path Tracer"
3. "Refraction Ray Cones"
4. "Many Lights With Grid-based Reservoirs"

[苏良 @leonsu123]

发现一本让人大开眼界的书，A thousand brains.

从没见过一本书能把学习的本质讲的如此透彻的，相见恨晚。

[Dexter Yang ᯅノ🌐🔗🧙🏻‍♂️👾🦋 @dexteryy]

Orion（Meta 的终极形态 AR 眼镜）信息汇总

三种交互方式：Vision Pro 模式的眼手交互、雷朋眼镜模式的语音交互、基于神经接口（EMG）的手势识别+眼动

支持多任务空间应用和虚拟与现实融合，FOV 达到 70 度，重量 98g，用镁制镜框散热，7 个摄像头（眼动跟踪镜头在镜框两侧，手势和房间跟踪的摄像头
在支架上）

眼动和手势跟踪、SLAM、AR world locking 都在眼镜上运行，十几个定制芯片让这些算法的功耗从几百毫瓦缩减到几十毫瓦

应用逻辑在 wireless compute puck 上运行，不能距离超过 12 英尺。puck 上有摄像头和触控板（好像最初是准备用来做控制器，后来被手部追踪和 EMG 取代）

EMG 手环支持的手势有：���指捏食指确认选择，拇指捏中指是home键，拇指往外弹是上下滚动

FOV 是 Orion 的圣杯，靠碳化硅镜片实现，导致22年时成本高达 $10000，只做了 1000 副用于测试和演示

碳化硅镜片的 FOV 大但分辨率低，目前 Orion 有两个版本：每度 13 像素和每度 26 像素

计划未来几年发布的第一个消费者版本是 Orion 二代，不用碳化硅镜片，FOV 更小分辨率更高，预期价格相当于高端手机和笔记本（$1000）

[Halim Alrasihi @HalimAlrasihi]

This is Hollywood grade lip-syncing:

Here is how to create an AI avatar with accurate and realistic Lip-Sync. This might be currently the best tool for video-to-video right now.

Only 4 steps are needed:

[小互 @xiaohu]

多邻国推出 AI 视频通话功能

你可以直接通过视频通话来学习语言和练习口语

感觉是接入了OpenAI的 高级语音API功能，这意味着高级语音版本的API已经给了一些合作开发者开发产品了。

估计API也将很快发布！

[el.cine @EHuanglu]

MIMO: Advanced Video-to-Video Model by Alibaba

This is insane!

With this v2v model, we can take a video of ourselves and use it to animate any character we create, all while keeping the original background from the video.

It's like we're putting ourselves into any other new worlds!

6 Example:

[Tom Huang @tuturetom]

卡内基梅隆大学 @UsePenrose 开源的这个 6.8K Star 🌟交互式图表库也太好了！！⚡️ 感觉和 3brown1blue 的 manim 有的一拼🤔

只需以纯文本输入数学符号即可创建漂亮的图表！👨‍💼 所有的坐标均有布局引擎自动计算，并且可以直接重用为 HTML/CSS 等前端组件🔥

开源地址 👉 https://t.co/KVwhbI9RbM

[Runware @runware]

Want stunning image generation in the blink of an eye? With https://t.co/xrZ3Y7k6Q5, create accurate, high-quality images in under 1 second—completely free to use! 🎨

[Thinking Garden @chenqing663]

《Hands-On Large Language Models》在本周读完了，这是今年唯二的我读过的关于大模型的最棒的书，另一本是《Build a Large Language Model (From Scratch)》。
---
当然也是买的最贵的两本电子书😂。

[BigguyDev @BigGuyCode]

Working on structural connectivity system. Arrows show direction of load bearing connections, color shows load ratio (red means high degree of stress). This system will allow buildings to collapse under their own weight and be destroyed in more interesting ways
#gamedev #indiedev

[Khronos Group @thekhronosgroup]

Once @Microsoft 's Shader Model 7 is released, #DirectX 12 will accept shaders compiled to SPIR-V. Microsoft is working with the Khronos #SPIR and #Vulkan Working Groups to ensure that this transition benefits the whole development ecosystem.

https://t.co/QA5ogxxyqN

Microsoft

[Math Cafe @Riazi_Cafe_en]

NASA's "Introduction to Tensors": https://t.co/TxdhuFRMiG

[AIGCLINK @aigclink]

斯坦福大学开源的一款研究想法生成器：AI-Researcher，直接将提案效率拉满！

可以将想法转化成详细的项目提案，并评估提案的可行性，AI-Researcher给出的提案足够详细，甚至于可以直接按照提案中的步骤执行整个项目

特点：
1、通过一系列模块，相关论文搜索、基于语境的想法生成、想法去重、项目提案生成、项目提案排名和项目提案过滤，自动化生成高质量的研究想法

2、这些模块被设计为顺序运行的端到端创意生成流程，也可以作为独立的研究辅助工具单独运行

3、可以根据不同的研究领域和需求进行调整，比如可以通过添加新模块来扩展其功能

github：https://t.co/V41PrLmbC7

#AI研究助手 #AI提案助手

[小互 @xiaohu]

Fish Speech 1.4 发布

完美支持中英日8种语言的开源TTS模型

- 70 万小时的多语言数据进行训练（高于 20 万小时）

- 支持 8 种语言：英语、中文、德语、日语、法语、西班牙语、韩语和阿拉伯语

- 完全开源

- 超快的 TTS 和超低延迟

- 即时语音克隆

- 支持自托管或云服务部署

- 提供 API 服务

官方网站： https://t.co/XUWZga6Yvn

GitHub: https://t.co/AOLRpJYQtB

HuggingFace : https://t.co/51H61rGR6X

Demo: https://t.co/UkmpYk56Cm

[Kay 🎨 @KayLousberg]

platformer game assets
#gamedev #indiedev

[Eric @EricLeeKK]

吴恩达的AI+Python编程课，11节课，每节课5分钟左右。1���时候教会编程。他做到了！牛逼！
https://t.co/8owVtgwio9

[c0de517e/AngeloPesce @kenpex]

It's been a while that I haven't posted this https://t.co/MHjSpFPjc1

[Rune Skovbo Johansen →🦋🐘 @runevision]

I've rewritten the signed distance field generator by @catlikecoding to be 40x to 75x faster by using Unity Burst. It's available here (licensed partly MIT and partly MPL 2.0):
https://t.co/mUrxGhFiH8