~100 days ago Grok 4.5 came out and I thought it was a massive breakthrough. Today I'm running Qwen3.8-Flash-Next at Q6 on 10 year old hardware at ~40 tok/s, and on the Artificial Analysis Intelligence Index it scores 40 vs Grok 4.5 (high) at 39. Same chart attached. This is moving so fast.
绝了!NVIDIA最新开源了2,575 个高规格 3D 资产,整整472GB!!
从日常用品到工业设备,里面收录了不少现实环境里的物体。
一揽子,全部放进了 Hugging Face。
NVIDIA SimReady Assets。
绝不单纯是一个“免费建模素材库”!!
NVIDIA给每个3D 资产都带上了物理属性和语义信息,比如有多重,什么材质,重心在哪,用在什么场景等等,全都写进了参数里。
明摆着,就是用来让AI学习的!
学习怎么和真实的物理世界互动!
传统的3D资产在AI眼中,就是个有贴图的“视觉资产”,NVIDIA想做的是,把“视觉资产”变成AI可以理解的“物理资产”!
所以,这批资产的价值可想而知!
至少能够用到这些场景上!
机器人抓取和搬运训练,先在虚拟环境里练习拿取、摆放物体;
搭建工厂、仓库的数字孪生,测试设备和物体的布局;
配合合成数据工具,改变光照、视角、遮挡和物体位置,生成带标注的训练数据。
大家都在卷文字、代码与图像生成时,NVIDIA 开始为Physical AI(物理 AI) 铺设地基了。
不过,对大部分人来说,也是一个相当全面的3D资产素材库!
链接在下面,必须收藏了!
llm-visuals 0.14 is out. Watch your local LLM think, live in the terminal. New this time: alerts that speak up once when a card fills up or runs hot, an expert balance view that shows whether the MoE router is spreading work or collapsing onto a few experts, and a memory pipeline view that shows where the tokens wait. Free and open source. What should it show next?
Local LLM hoarders, this one's for you. model-space scans your box for every GGUF, safetensors and Ollama model, shows what's eating each disk, and moves or deletes them without breaking your services. Mine found 96 models and 2.6 TiB. Free and open source, link below.
Kookabus crew episode 5: the I2C bus. Wrong address, NACK. Bridge A0, 0x14 + 1 = 0x15, ACK. Every board on the bus needs its own address, and Kookabus-02 gives you 0x14 to 0x17 with two solder pads. Next up, the servo showdown. Which board should get an episode after that?
Your AI PCB designer doesn't need to live on the same machine anymore. WholeBoard now searches your whole LAN for local model servers and lets you pick one with a click. Still free and 100% open source, and the code's in the repo below.
Mac folks, the open-source KiCad autorouter now builds on macOS 15 and 26 with a Metal backend (huge thanks to Chris Hemingway). TraceMaker 0.11 also adds a run report that explains every connection, plus a new docs site. 177/177 on a Teensy audio board. What should I route next?
Who needs an internet connection to have AI design PCBs? The open-source WholeBoard now searches for any local models running and uses them. Ollama, LM Studio, llama.cpp, vLLM, Jan, Strata. One click and it's designing boards at $0 per million tokens. What are you running locally?
7 hours until my weekly reset, 45% of my credits left. So I challenged Claude Opus 5.5 to build a full rally game from my childhood memories. Pacenotes, splits, a ghost car, and a co-driver who sounds like a robot. Which game would you rebuild?
WholeBoard now sims the board inside its enclosure. ESP32-C3 board, 333 mW: 42.6 °C in open air, 52.0 °C sealed in a PLA box. Add lid and side slots and it drops to 50.4 °C. Still in spec in a 40 °C room. Sealed or vented for yours?
github.com/DingoOz/WholeBoar…



