From 32 items, 3 important content pieces were selected
- Apple SHARP runs in-browser with ONNX Runtime Web ⭐️ 8.0/10
- NASA’s Artemis II laser link sends 484 GB from lunar vicinity ⭐️ 8.0/10
- DeepSeek-V4 Preview Goes Open Source ⭐️ 8.0/10
Apple SHARP runs in-browser with ONNX Runtime Web ⭐️ 8.0/10
A developer exported Apple’s SHARP single-image 3D Gaussian splatting predictor to ONNX and got it running entirely in the browser with onnxruntime-web and the WebGPU execution provider. The demo lets users upload an image, generate a .ply 3D output, and preview or download it locally without sending the image to a server. This is a strong example of browser-side machine learning becoming practical for heavy vision models, not just lightweight demos. Running SHARP locally improves privacy and removes the server round trip, which could matter for creative tools, edge devices, and future client-side AI workflows. The author says the exported model is large, with a roughly 2.4 GB sidecar, so the first load is slow on a cold cache even though inference takes only a few seconds on a recent Mac. Apple’s released weights are marked research-use only, and while the demo hosts the ONNX export on R2 for convenience, users can also export their own from the upstream Apple repository.
hackernews · bring-shrubbery · May 3, 09:14
Background: SHARP is Apple’s recent monocular 3D reconstruction model that turns a single image into 3D Gaussian splats, which can be rendered as a scene. ONNX is a portable model format, and ONNX Runtime Web can execute those models in the browser; the WebGPU execution provider lets the runtime use the client device’s GPU for more demanding workloads. 3D Gaussian splatting is a representation for visualizing and rendering 3D scenes from learned point-like primitives rather than traditional meshes.
References
Discussion: Commenters were broadly impressed by the size and technical ambition of the ONNX export, with one noting that a 2.4 GB ONNX file is remarkable. Several people connected the demo to broader browser-native AI trends, especially privacy-preserving client-side inference and creative applications, while also pointing out caveats around WebGPU compatibility, model conversion friction, and the practical limits of very large models in the browser.
Tags: #ONNX Runtime Web, #WebGPU, #3D Gaussian Splatting, #Browser ML, #Computer Vision
NASA’s Artemis II laser link sends 484 GB from lunar vicinity ⭐️ 8.0/10
NASA’s Artemis II optical communications system, O2O, successfully downlinked 484 GB of data from the lunar mission at up to 260 Mbps. The transmission showed that the laser link can move high volumes of mission data from the Moon’s vicinity back to Earth far faster than traditional radio links. This is an important validation of high-bandwidth optical communications for future lunar and Mars missions, where crews and spacecraft will need to send back more video, images, and science data. It could improve real-time operations, enable smoother public-facing video, and reduce the communications bottleneck for deep-space exploration. The O2O module was developed by MIT Lincoln Laboratory and uses a 4-inch telescope with two gimbals to point laser communications toward Earth ground terminals. NASA’s ground infrastructure for the demo included JPL, White Sands, and the Australian National University’s Stromlo Observatory, and the system reportedly received 26 GB in under an hour at one point.
telegram · zaihuapd · May 3, 00:50
Background: Optical communications use infrared or laser beams instead of radio frequency waves to send data, which can significantly increase bandwidth. NASA’s Artemis II is a crewed lunar mission, and O2O is intended to demonstrate that lasercom can return HD video and high-resolution images from deep space. The technology is part of NASA’s broader push to build communications systems that can support more demanding missions beyond Earth orbit.
References
Tags: #NASA, #laser communication, #Artemis II, #space systems, #deep-space communications
DeepSeek-V4 Preview Goes Open Source ⭐️ 8.0/10
DeepSeek has released the preview version of DeepSeek-V4 and made it open source. The DeepSeek-V4-Pro variant is reported to significantly improve agent capabilities, while DeepSeek-V4-Flash offers a smaller, cheaper option for API use. This matters because it strengthens the open-source frontier for agentic AI, especially in coding and reasoning tasks. If the benchmark claims hold up, more teams could adopt DeepSeek for production agents without relying only on closed models. According to the provided summary and search results, DeepSeek-V4-Pro is described as open-source SOTA for agentic coding and as outperforming publicly tested open-source models on math, STEM, and competitive coding. DeepSeek-V4-Flash is positioned as a lighter model with fewer parameters and lower activation cost, which should make inference faster and API pricing more economical.
telegram · zaihuapd · May 3, 02:21
Background: Agentic AI refers to models that can plan and carry out multi-step tasks more autonomously, such as coding workflows or tool use. In AI releases, benchmarks like math, STEM, and coding are often used to compare reasoning and practical problem-solving ability across models. Open-source model launches are especially important because they let developers inspect, run, and adapt the model themselves.
References
Tags: #DeepSeek, #open-source LLM, #agentic AI, #foundation models, #AI benchmarks