MattVidPro

I tested an opensource world model today and it’s one of those “this is clearly early… but also clearly the future” moments. Still janky, but insanely promising. Also checked out Google’s updated Stitch and AI Studio builder, and honestly the barrier to making websites/apps with AI keeps dropping fast. Google's habit of building random tools is paying off in the AI era. Check it: youtu.be/__LXXnB7ZBA

@KeyTryer Never played but I heard its about walking or something

what the fuck happens in this game

论文来了。名字叫 MSA，Memory Sparse Attention。 一句话说清楚它是什么： 让大模型原生拥有超长记忆。不是外挂检索，不是暴力扩窗口，而是把「记忆」直接长进了注意力机制里，端到端训练。 过去的方案为什么不行？ RAG 的本质是「开卷考试」。模型自己不记东西，全靠现场翻笔记。翻得准不准要看检索质量，翻得快不快要看数据量。一旦信息分散在几十份文档里、需要跨文档推理，就抓瞎了。 线性注意力和 KV 缓存的本质是「压缩记忆」。记是记了，但越压越糊，长了就丢。 MSA 的思路完全不同： → 不压缩，不外挂，而是让模型学会「挑重点看」 核心是一种可扩展的稀疏注意力架构，复杂度是线性的。记忆量翻 10 倍，计算成本不会指数爆炸。 → 模型知道「这段记忆来自哪、什么时候的」 用了一种叫 document-wise RoPE 的位置编码，让模型天然理解文档边界和时间顺序。 → 碎片化的信息也能串起来推理 Memory Interleaving 机制，让模型能在散落各处的记忆片段之间做多跳推理。不是只找到一条相关记录，而是把线索串成链。 结果呢？ · 从 16K 扩到 1 亿 token，精度衰减不到 9% · 4B 参数的 MSA 模型，在长上下文 benchmark 上打赢 235B 级别的顶级 RAG 系统 · 2 张 A800 就能跑 1 亿 token 推理。这不是实验室专属，这是创业公司买得起的成本。 说白了，以前的大模型是一个极度聪明但只有金鱼记忆的天才。MSA 想做的事情是，让它真正「记住」。 我们放 github 上了，算法的同学不容易，可以点颗星星支持一下。🌟👀🙏 github.com/EverMind-AI/MSA

Our AI Studio vibe coding roadmap for the new few weeks: - Design mode - Figma integration - Google Workspace integration - Better GitHub support - Planning mode - Immersive UI - Agents - Multiple chats per app - Simplified deploys - G1 support And more, should be fun : )

Composer 2 is now available in Cursor.

Today, we’re launching a new way to create with AI. With OpenArt Worlds, you can generate a fully navigable 3D environment from a single prompt or image, step inside it, and capture shots exactly the way you envision them. No more starting over. No more inconsistent scenes. You build the world once - and create inside it. • Move through your scene freely • Find your angles • Add characters and elements • Capture production-ready shots

We don’t generate videos. 🎬 We generate worlds from videos. 🌍 Introducing InSpatio-World — the world's first open-source real-time 4D world model‼️ Your input: a video clip Our output: a dynamic, navigable, persistent world 🕹️ explore freely across viewpoints ⏪ control time forward and backward 🔓 open-source and ready to build on :) Live demo: 🔗 world.inspatio.com Code & weights: 🔗 github.com/inspatio/inspa… Project page: 🔗 inspatio.github.io/inspatio-world

From Nvidia’s GTC, Jensen calls this “probably the single most important chart for future of AI factories”. Y-axis is “Throughput” (total volume) while X-axis is “Token Speed” (more tokens per second = more interactivity for a user + more context + more reasoning). Firms market and price token offerings on those two variables, which are in tension. A free tier typically is high throughput but lower token speed. Meanwhile, the priciest tier would have lower througput but high-value tokens (eg. research, coding) Nvidia’s challenge is to build systems that lift the entire line up and to the right. Jensen says Vera Rubin architecture improves revenue opportunity 5x vs. Blackwell. Then, if you add Groq to Vera Rubin, that revenue opportunity is up 10x vs. Blackwell (Groq useful for the higher value tokens).

Introducing the new @stitchbygoogle, Google’s vibe design platform that transforms natural language into high-fidelity designs in one seamless flow. 🎨Create with a smarter design agent: Describe a new business concept or app vision and see it take shape on an AI-native canvas. ⚡️ Iterate quickly: Stitch screens together into interactive prototypes and manage your brand with a portable design system. 🎤 Collaborate with voice: Use hands-free voice interactions to update layouts and explore new variations in real-time. Try it now (Age 18+ only. Currently available in English and in countries where Gemini is supported.) → stitch.withgoogle.com

Original vs DLLS 5 with Tweaked Tonemap vs DLLS 5 via: Reddit reddit.com/r/hardware/com…

I made a tool for vibe-coding real-time apps for your audience. think Slido/Kahoot but more meta. It can generate polls, quizzes, multiplayer games, chats basically anything collaborative that you can share during a talk or online meeting local-first and crazy fast thanks to @instant_db

CS:GO + Adobe + Wan2.1 = WorldCam. Interactive autoregressive 3D gaming worlds. > AI now generates playable 3D worlds live > You move the mouse, the AI builds the map instantly > Turn around, and it remembers exactly what was there No traditional game engines. Just neural networks hallucinating reality. cvlab-kaist.github.io/WorldCam/

FlashMotion. and it's Wan2.2-TI2V again. Few-step controllable video gen. - Precise multi-object box/mask guidance; - 50x speedup over SOTA; - supports multi-object tracking and camera motion. intuitive video editing tool. Easily animate static photos, direct exact movement paths for characters etc quanhaol.github.io/flashmotion-si…

Today we're starting to test an early version of our V8 model with our community. It's much better at following prompts, 5x faster, has native 2K modes, improved text rendering and the best personalization, sref, and moodboard performance ever. Have fun!

My preferred hardware stack, wheel in the back is for emergency steerability.

@javilopen I hear you, NB2 attempting to correct the image:

For MJ to still be putting out stuff with anatomical errors this bad in 2026 honestly feels a bit like throwing in the towel. If Nano Banana and the other frontier models didn't exist, I'd think we'd hit an AI plateau 😅 They still get a pass because OF COURSE, when it comes to aesthetics and artistic taste, they're still the best! It's kind of amazing how they've managed to hang on by that thread.

Midjourney v8 just came out, and I'm seriously disappointed. V7 left V8 right Seriously??? 😂

Tomorrow, we’re introducing you to your new vibe design partner. 🤝 Our biggest update ever drops tomorrow. 👀👇

We're shipping a new feature in Claude Cowork as a research preview that I'm excited about: Dispatch! One persistent conversation with Claude that runs on your computer. Message it from your phone. Come back to finished work. To try it out, download Claude Desktop, then pair your phone.

The frontier has increasingly shifted to hybrid models - from Qwen to Kimi-Linear and now with NVIDIA's Nemotron-3 Super - that rely on a strong linear sequence model. Today we release Mamba-3, the most powerful linear model to date. x.com/_albertgu/stat…