X / Twitter
PE3R: Perception-Efficient 3D Reconstruction. Take 2 - 3 photos with your phone, upload them, wait a few minutes, and then start exploring your 3D world via text!
https://t.co/0r407Hky5i https://t.co/6VR5o0FYLj
Introducing Chapter-Llama [#CVPR2025], a framework for 𝐯𝐢𝐝𝐞𝐨 𝐜𝐡𝐚𝐩𝐭𝐞𝐫𝐢𝐧𝐠 using Large Language Models! 🎬🦙
Check it out:
📄 Paper: https://t.co/1KhPsgZYUN
🔗 Project: https://t.co/68GevYyznx
💻 Code: https://t.co/MysWVlewRm
🤗 Demo: https://t.co/zKmL6v3PKU https://t.co/P0V0PWG4uq
are:na · mobile client
by @_yihui https://t.co/buSz3xaprQ
You can just build MCP servers for platforms you use if there's an API. No need to wait.
Here's the example of building a spiral-MCP (@every), showing the steps from docs to tested tools: 🧵 https://t.co/CRMYTRTl6Q
2025 CSS Reset https://t.co/g5V6UVYmqZ https://t.co/hs9zQLif0c
Today, I'm excited to launch Arrakis: an open-source and self-hostable sandboxing service designed to let AI Agents execute code and operate a GUI securely.
GitHub: https://t.co/BXuzxWEJxE
Watch Claude code a live Google Docs clone using Arrakis. Having a VM sets it free -🧵 https://t.co/vE2WqUUSvI
Sound on for this one.
We wired up @OpenAI's new STT and TTS models into a single voice agent. The results are super fun! Link to a live demo in the next tweet. https://t.co/mVfwPGlbxP
🚀 Vapi 2.0 is live on Product Hunt. A big milestone for the team!
Vapi is the infrastructure layer for real-time voice agents.
It lets developers deploy LLM-powered phone agents that can talk, listen, reason, and act with sub-500ms latency and support for 60+ minute conversations.
What’s new in 2.0:
- 100+ model integrations (transcription, LLM, TTS)
- Conversation workflows and knowledge bases
- Vapi voices that sound human
- Rock-solid enterprise infrastructure to scale to millions of calls
- Real-time observability with Langfuse integration
- A metric ton of voice agent configuration options
We’ve now powered over 44 million calls across healthcare, support, logistics, and more.
Try it live: 1-844-HEY-VAPI
Show us some love: https://t.co/H94LUaYqzy
PS: Team Vapi has hidden Easter eggs on their Product Hunt page 👀
Find them and DM us what you spot to win a slice of the $10,000 prize pool!
Using Memory Bank locally is powerful for giving Cline persistent context for your individual workflow. But engineering teams face a bigger challenge: shared context and knowledge silos.
We saw a startup tackle this brilliantly. They adapted the Memory Bank concept by swapping local files for Notion pages. Now, their Cline instances use a Notion MCP server to automatically read and update this shared knowledge base.
The implications? Team context syncs automatically. Documentation becomes a living system updated by the AI. Knowledge silos disappear. Everyone – developers and AI – operates from the same, consistent, collective intelligence. It's a step towards truly integrated team + AI workflows.
Wow! OpenHands LM is now pretty easy to use locally thanks to @lmstudio! https://t.co/5ecZysgIze
Thrilled to announce our paper “Dynamic Concepts Personalization from Single Videos” is conditionally accepted to #SIGGRAPH! Unlike static concepts, our method captures dynamic concepts, enabling advanced editable compositions for the first time! https://t.co/L10kWZaiJx https://t.co/kFnVKWp5IE
Imagine Cline autonomously navigating websites, clicking buttons, and extracting data. The new Stagehand MCP server from @browserbase makes it possible, using AI to drive a real browser for complex web automation tasks.
The Stagehand MCP: now available in the MCP Marketplace. https://t.co/hDPm4Jf3UY
Today, we're excited to make two big announcements!
- OpenHands LM: The strongest 32B coding agent model, resolving 37.4% of issues on SWE-bench Verified 📈
- OpenHands Cloud: SOTA open-source coding agents from your computer, phone, github, with $50 in free credits 🙌☁️ https://t.co/xvHzs7BWhQ
works for prompting https://t.co/F6qQz4tvAK
Video generation is getting increasingly 🎨 visually appealing, but what’s next? e.g., Do these models truly follow 🌍 real-world principles?
We introduce 𝗩𝗕𝗲𝗻𝗰𝗵-𝟮.𝟬: a next-gen benchmark for evaluating 𝗶𝗻𝘁𝗿𝗶𝗻𝘀𝗶𝗰 𝗳𝗮𝗶𝘁𝗵𝗳𝘂𝗹𝗻𝗲𝘀𝘀
🔬 Physics, 🤔 Commonsense, 🦴 Human Fidelity, 🎭 Creativity, 🎮 Complex Controllability
✅ 18 fine-grained dimensions
✅ Fully automatic & open-source
✅ Human-aligned via large-scale validation
#AI #VideoGeneration #VBench
Search is the natural abstraction for augmenting AI with moving context
The initial excitement around RAG centered on vector databases. Since then, we've seen various variations—graphRAG or lazyRAG—each offering a spin tied to a specific technology or vendor.
A slight digression: developers love to connect a method with a specific vendor. I see it all the time in my mentions on social media. We use X for Y. MongoDB defined the NoSQL database paradigm. In 2025, there’s no real technical reason to use it, but the stickiness remains. Marketing to developers works.
Back to the RAG concept. The closest and most durable abstraction is search—not just in the traditional “10 blue links” sense, but as a broader mechanism for finding relevant information (context) to help a model perform a task.
That task could be answering a question (as in the original RAG paper), but it could just as well be refactoring code or drafting a legal proposal. The kinds of tasks we want AI to help with are endless—and that the quality of those tasks hinges on search is, to me, undoubtedly true.
If that search is implemented with a dense or a sparse vector or involves hybrid or whatever, that is more of an implementation detail. The level of abstraction is searching.
Here’s a glimpse of what that looks like in practice. These are the tools available to Windsurf (AI-assisted coding, one of the hottest applications of AI).
Do you notice anything in the screenshot? Many of the available functions—what we now call tools—are search tools.
Search is the natural abstraction for augmenting AI with moving context.
We are releasing OpenDeepSearch (ODS), an open-source search agent that works with any LLM. When paired with DeepSeek-R1, ODS outperforms OpenAI’s specialized model for web search, GPT-4o-Search, on the challenging, multi-hop FRAMES benchmark from DeepMind (+9.7% accuracy). https://t.co/fsKrenbZqC