Meta Max is a 5-foot, 106-lb humanoid robot prototype built in three months. It runs on NVIDIA Jetson Thor, has expressive eyes, a mouth, and five-fingered hands that can pick up small objects.
Key points
Max 2 design includes improved hands, added stair-navigation sensors, and a built-in projector.
Meta is hiring robotics engineers in China to scale development toward mass production.
Sources
Neuron Studio runs Qwen3.5-4B in browser, 2.91GB download
Neuron Studio runs Qwen3.5-4B directly in the browser, letting users select layers and tweak intensities with side-by-side comparison. The model is about 2.91GB and must download fully the first time; no server or API is required.
Key points
Model: Qwen3.5-4B runs in the browser.
Size: The model is about 2.91GB and downloads fully on first use.
Sources
- Original post
- huggingface.co - Linked in the post
Halloween Special Stream Update – Live on Roku and AppleTV via AlchemyStream
Halloween Special is live now on Roku hardware devices – search "AI Music Video Show". It is also available on Roku plus AppleTV via AlchemyStream. X will not go live until glitches are fixed.
Key points
Live on Roku hardware devices; search "AI Music Video Show"
Available on Roku + AppleTV via AlchemyStream
Sources
- Original post
- musicvideochannel.ai - Linked in the post
Junfan Halloween Robotics Building Maker Arm for Dream Lab's Dream-1 10/31
Junfan Halloween Robotics will build a Maker Arm for Dream Lab's Dream-1 on 10/31. The event includes a keynote by Ryan_Resolution (@makermodsai). World action models are opening new possibilities for language-guided manipulation.
Key points
Event: Junfan Halloween Robotics building Maker Arm for Dream Lab's Dream-1 on 10/31.
Keynote speaker: Ryan_Resolution (@makermodsai).
Topic: World action models are opening new possibilities for language-guided manipulation.
Sources
Claude Bootcamp Day 1 – two-day build at ONBTC with Claude API
Claude Bootcamp Day 1 announces a two-day building session at ONBTC using Anthropic’s Claude API. The session is led by Moisés Iglesias and Rodrigo Trujillo.
Key points
Event: Two days at ONBTC building with Anthropic’s Claude API.
Leaders: Moisés Iglesias (@JMoises) and Rodrigo Trujillo (@Trujasx).
Sources
CoreWeave AI engineers work with Aston Martin F1
CoreWeave sent Morgan McGuire and Agata Mlynarczyk to work with Aston Martin F1, merging generalist ML skills with a specialized engineering environment.
Key points
Sent: Morgan McGuire and Agata Mlynarczyk
Collaboration: CoreWeave with Aston Martin F1
Sources
OpenJaws Open Hardware UMI Framework with Stereo Localization released
OpenJaws is an open hardware and software framework for sensorized Universal Manipulation Interface that uses a modular 3D-printed gripper and low-cost stereo camera to obtain metric camera trajectories. It introduces the UMI-Loc-Bench-v1 benchmark and shows stereo localization outperforms monocular alternatives. Code, data and videos are released under a commercially permissive license.
Key points
OpenJaws pairs a modular 3D-printed gripper with a low-cost stereo camera for metric camera trajectories.
UMI-Loc-Bench-v1 benchmark quantifies real-world localization accuracy for handheld data-collection devices.
Sources
- Original post
- OpenJaws: An Open Hardware UMI Framework with Stereo Localization for Generalist - Linked page
- open-jaws.github.io - Linked in the post
Nonlinear Direct Feedback Alignment for Scalable Backpropagation-Free Training announced
Nonlinear Direct Feedback Alignment (NDFA) introduces a lightweight nonlinear surrogate subnetwork for each hidden layer’s feedback. The paper reports that NDFA consistently outperforms existing DFA methods and narrows the performance gap to backpropagation across CIFAR-10, CIFAR-100, SVHN, and ImageNet. Code is released alongside the arXiv submission.
Key points
Introduces NDFA, using a surrogate subnetwork to construct nonlinear feedback for each hidden layer.
Experiments on residual networks and vision transformers show NDFA beats prior DFA approaches on multiple vision datasets.
