AI Agent • AI Infra • AI Game https://xitter.cf/t.co/tV2peE5n0u - Artifacts for every harness https://xitter.cf/t.co/hFgLIi52fa - TUI agent dashboard Post in 中文/English
Joined February 2024
- Tweets433
- Following183
- Followers195
- Likes119
Pinned Tweet
r3.hyperlogue.tech the r3 product site just launched together with the latest r3 v2.2.0.
r3 is like Claude Artifacts, but focusing on viewing & feedback. It works with any agent harness.
I use r3 in almost all my coding sessions. In the last week, agents created 32 artifacts, published 102 versions, handled 133 feedback threads with 300+ comments.
The use cases include design review, ui prototype, pre-launch changelong revise, launch video review, code base explanation. It's so versatile and deeply integrated into my agentic workflow.
I'd love to hear what you love about claude artifacts or chatgpt site. We can improve r3 to make the features you love available to everyone on every harness!
想到的视频是这个 xitter.cf/cheerselflin/status/21…
Today I learned about shadow DOM.
It’s for attaching a self-contained DOM tree to an element. CSS is isolated between the page and this shadow DOM.
Why does it matter? I use this technique to put dozens of UI mockup frames, each built from my project’s real UI components, onto one single html page. Review UI variants side by side, with no clashing classes or styles. Events and animations keep working inside each frame.
r3 v2.4.0 has better support for shadow DOM. In comment mode, you can select elements inside shadow DOM, like a button inside one of the mockup frame, and add comment on it. This gives your agent the perfect context to take action on your comment. You can just comment ‘make it 6px longer’, rather than ‘make the confirmation button 6px longer. It’s under the text input on the model setup page.’
Used it to review the full UI for a demo we're about to launch, all in one pass. Super smooth.
Alternatively you can use iframe. It provides better isolation at the cost of performance. In my experience iframe is too slow if you have more than 10 frames
试了一下 openflim 来做 demo 视频，感觉很好用。之前让 opus 直接做也行，但如果要改很多细节，这种纯对话的方式就不方便了。
openfilm 提供了一个网页版的编辑器，可以手动调整，做一些微调的时候会很方便。
github.com/openfilm/openfilm
r3 在最新版里提供了中文界面，文档站也有中文了。喜欢用中文界面的小伙伴可以试一试。
r3 是一个类似 Claude Artifacts 的工具。它专注于评论功能，让你能和 agent 更好地在多轮对话中追踪需要关注的问题。r3 没有做多用户分享或者公开访问之类的功能。
r3 完全在本地运行，并且能适配任何 agent harness。
第一版中文文本是由 gemini 3.8 从英文翻译的。虽然是大型 LLM，翻译出来的结果还是有机翻的味道😂
我自己审阅了一部分，整个审阅的过程也是让 agent 直接写了个 html 放在 r3 里让我提意见，算是 dogfooding 了。
完全用中文来说这些技术词汇的确是有难度。比如 Artifact 这个词我就很难找到合适的翻译，“Review captured changes with the original lines and context intact.” 这句话完全不知道怎么翻译。所以 r3 的中文界面可能会读起来很生硬，希望小伙伴们能多给反馈。
Me: seeing the nodejs rce. PANIC. jump up from couch and start to patch my servers.
5 minutes later: wait… we run Bun not node. We are safe.
5 minutes later: wait… does Bun use the same brotli library? Yes?? And it doesn’t have a patch release yet?
Two UNO reverse cards in ten minutes, and I'm back to panicking.
Small difference between Grok Bot and ChatGPT Dots in voice mode:
Say goodbye to Dots: it says goodbye, then stays on the line.
Say goodbye to Grok Bot: it says goodbye and hangs up.
Tiny detail, but a very clear win for Grok Bot. What do you think?
Amazing news! There must be so much to learn from such a specialized inference stack. They have released their model, kernels and now the inference engine. I wonder what they will open source next? Maybe the training part of the stack?
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/DeepI…
1/4
Rick Ye retweeted
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/DeepI…
1/4
Rick Ye retweeted
Today, we're open-sourcing 5 Ming-3.0 models across 3 speech families.
- Ming-3.0-tts: 0.5B voice design & paralinguistic cues; Tiny 7.9B A1.3B long-form speech.
- Ming-3.0-audio: Flash 126B A6.5B understanding & tools; Tiny 9.3B A2.7B efficient dialogue.
- Ming-3.0-asr: Tiny 8.6B A2B dialects & context.
Rick Ye retweeted
🦆🔫
He convertido un AirPod en un mando para jugar al Duck Hunt.
Lo giras para apuntar y das un toque en el micrófono para disparar.
Sin tocar el ordenador, full vibecodeado por voz al mientras dormía a la bebé 😅
Changed borders to green for successful action notifications in captain-miao.
The agent agreed it's better UX. Wondering why didn't it do this on the first pass? Did my prompt over-constrain its taste, or leave too much slack?
How can we get good UI from AI in the first shot?
I've been using Grok Bot and ChatGPT Dots heavily lately, even talking to them in voice mode while I am running. Both have been huge productivity boosters, but each shines in very different ways:
Grok Bot
- Multi-bot workflow: Running multiple bots keeps tasks organized and contexts cleanly isolated.
- Group chat: looks promising in theory, but in practice it gets too chatty without noticeably improving the outcome. I prefer having one bot pinging another bot to transfer context.
- Voice mode: Built for pure efficiency. Well-designed product-wise. It acknowledges my request in a few words, executes, reports back. After hanging up, it drops a text recap with the full transcript available in a few clicks.
ChatGPT Dots
- Long-running tasks: Rock-solid for long tasks. I use it primarily to generate 3D models via Blender. It can reliably work overnight.
- gpt-live: The quality is really really impressive. It genuinely feels like talking to a real person. However I don't find a way to see the transcript, and no recap is given after the call.
How has your experience been with these two products? Which one are you finding more useful day-to-day?