Robin Delta (@heyrobinai)

china is not slowing down and this one is insane.. Chinese researchers just open-sourced SLAM3R, it rebuilds a full 3D scene from a single regular video, in real time. no LiDAR, no depth sensor, no camera rig. one phone-grade video in, a 3D world out. > runs at 20+ FPS on a single GPU > monocular RGB, just walk around filming > feedforward neural net, no slow pose-optimization step > beats optimization-based methods on reconstruction benchmarks > handles long walkthroughs without drifting apart what makes this work: older methods solve for camera position frame by frame, which is why they're slow. SLAM3R skips that and regresses the 3D geometry directly. that's how it hits real time. 100% free and open source:

NOBODY IS TALKING ABOUT THE MOST EXPENSIVE STEP IN EVERY AI PIPELINE.. It isn't the model. It isn't the embeddings. It's PDF extraction and most teams are still feeding their LLMs flattened garbage. @nutrientdocs CLI quietly fixed this: → Messy PDF in, clean structured markdown out. Headings, tables, lists preserved, not mangled. → Runs locally. No signup, no API key. Free up to 1,000 docs/month. → One command: npm install -g @ pspdfkit/pdf-to-markdown. Point it at a file or a whole folder. → On Claude Code or Cursor? Install it as a skill and your agent calls it automatically when it sees a PDF. Your retrieval is only as good as your chunks. Your chunks are only as good as your extraction. Everyone is optimizing the top of the stack. The real unlock is at the bottom.