Peter Rigole

@elonmusk When can we expect Grok-code tools? Claude-code and GPT-codex are incredible software development agents, but we’re still waiting for the Grok alternative…

Grok 4.20 Heavy (Beta 2) is extremely fast for deep analysis. Beta 3 will have many fixes and functionality gains.

@giswqs This looks really nice! Can we also use standard background map layers, like we do in OpenLayers (openlayers.org) ?

Excited to share a new open-source project for the web-mapping and LiDAR community: maplibre-gl-lidar 🚀 With maplibre-gl-lidar, you can visualize large-scale LiDAR point clouds (LAS/LAZ, including COPC) directly in the browser using MapLibre GL JS. It supports point-cloud data stored locally or hosted on the web, and comes with an interactive GUI featuring customizable color schemes, classification legends, and point picking. It even works seamlessly inside Jupyter Notebooks. Visualizing tens of millions of LiDAR points has never been this simple or fast, making it easy to build modern, lightweight, and fully interactive web-based LiDAR viewers. Try it out: Live Demo Viewer: lidar-viewer.gishub.org GitHub Repository: github.com/opengeos/mapli… Full Video Tutorial: youtu.be/lV3oKk2WTEo #MapLibre #LiDAR #WebGIS #OpenSource #PointCloud #3DMapping

💯

@elonmusk Is a Grok version planned for interactive coding? As an alternative for Claude Code and GPT Codex?

Grok 🥇

@zhaohexu2001 Impressive work! Are you considering to add an open source license?

City-scale 3DGS 🏙️ doesn‘t need a GPU cluster Introducing CLM, a training system for large-scale 3DGS trains 100M+ Gaussian splats on single RTX 4090⚡️ (6x larger than GPU-only while keeping up to 90% throughput) 🔗 Homepage + Code are released! 📄 Accepted to ASPLOS 2026 🧵

@bingyikang Impressive work! Thanks for making this open source 🙏 Do you provide a streaming API in your code as well? I played around with it, but I was limited by the number of frames I could process on a 24 GB GPU

Spatial perception from multiple cameras Given several images of different viewpoints from a vehicle, DA3 estimates stable and fusible depth maps, enhancing autonomous vehicles' environmental understanding. We believe this is also achievable in robotics. 👇(5/n)

After a year of team work, we're thrilled to introduce Depth Anything 3 (DA3)! 🚀 Aiming for human-like spatial perception, DA3 extends monocular depth estimation to any-view scenarios, including single images, multi-view images, and video. In pursuit of minimal modeling, DA3 reveals two key insights: 💎 A plain transformer (e.g., vanilla DINO) is enough. No specialized architecture. ✨ A single depth-ray representation is enough. No complex 3D tasks. Three series of models have been released: the main DA3 series, a monocular metric estimation series, and a monocular depth estimation series. The core team members, aside from me: @HaotongLin, Sili Chen, Jun Hao Liew, @donydchen. 👇(1/n) #DepthAnything3

Let's fine-tune OpenAI gpt-oss (100% locally):

gpt-oss @openai is here! 🔥 > two MoEs with 21B/3.6B and 117B/5.1B total/active params, efficient reasoning models 🤯 > use & fine-tune with transformers & TRL 🛠️ > inference powered by @huggingface Inference Providers 🫡 > apache 2.0 license 💗

@_avichawla Can you do this on a few RTX3090’s?

5️⃣ Use GRPO and start training We employ GRPO, an RL method, to enhance reasoning, and start training! The setup includes: - Training config (LR, optimizer, steps) - Reward functions - Trainer execution Check this code and training logs 👇

Let's build our own reasoning LLM using Reinforcement fine-tuning (100% local):

In Europe, too many leaders are used to speak and say nothing. Conversations are not real because they have to be polite first instead of having content. This is why they can’t deal with what @JDVance said, simply because it is the truth. The truth is that the US Vice President made people angry in Munich by simply telling the truth, by engaging in a real conversation instead of saying what the elites want to hear. Indeed, European leaders are ignoring the will of their people, ignoring religious freedoms and not acting to halt illegal migration or protect the European values. That is not fake. Europe isn’t investing in itself. But indeed it is uncomfortable. You can’t change anything if you skip the uncomfortable facts and uncomfortable talks. When certain leaders stop being angry when facing the truth perhaps they will be ready for the change that we all need. Perhaps European leaders should have just replied honestly. No change without honest conversations. Honest conversations! Isn’t this exactly what the Munich Security Conference is for? Assita Kanko

@ValterMazzola @calebfahlgren It is available on ollama, you can pull the model and run it: ollama.com/library/qwen2.5

@calebfahlgren Mac model?

Qwen-2.5 on WebGPU 🏎️ • 42 tok/sec for Qwen2.5-Coder-1.5B on Mac ⚡ • Powered by MLC WebLLM and WebGPU 🔥 Watch Qwen2.5-Coder-1.5B build a website entirely in the browser!

@NM_Reid @BrentBrewington Either use virtualenv, or run it in docker containers if the dependency tree is complex. Tools like vscode directly attach to these containers and you can even run them remote on cuda enabled machines if you don’t have a GPU locally. Containers are trivial to rebuild anywhere.

@BrentBrewington git repos for academic software are often like "install our software and dependencies from source by spending three days in hell, or just do `conda install <pkg>`. " miniconda has been most folks default.

uhh, anaconda just sent a message to our HPC admins that we're in violation their ToS and we now need to pay for a license or remove all their software from our system?

so the local LLM action begin. Apple is leading the pack with the MLX work that allows you to run LLM’s locally. everything points towards private personal LLM’s at the WWDC in June…

#PyTorch goes to the Linux Foundation. This is massive for open source software! 👌🔥🎉 “As of August 2022, PyTorch was one of the five-fastest growing open source software communities in the world alongside the Linux kernel and Kubernetes.” linuxfoundation.org/blog/welcoming…

I am developing and teaching a new course on 3D Vision at Berkeley this semester: CS294: Geometry and Learning for 3D Vision. It will be finished in about two weeks. I will make all the lecture slides available. It is time to close the loop between reconstruction and recognition.

Sneak peek van een testopstelling voor het Surv-AI-llance project waar we met heel veel trots aan meewerken! Meer weten? Lees het volledige project hier: surv-ai-llance.com.

Today we installed the camera-radar setup at Corda Campus for the Surv-AI-llance project

Welcome to the future. Cop pulls over driverless car (because no lights?) Then Cruise goes on the lamb. (via instagram.com/tv/Cb1q-vggFn9/)

@steven_odb Idem hier, afzender contacteren en opnieuw laten versturen 😉

@alexeyab84 Looks impressive, how did you visualise the 3d point cloud?

DPT (Dense Prediction Transformers) State of the art 3D-reconstruction from a single RGB-image in Real-time (>30 FPS) with high Zero-Shot Accuracy on images different from the training ones. Paper ICCV2021: openaccess.thecvf.com/content/ICCV20… Code: github.com/isl-org/DPT RGB->Depth->3D