LLM-orchestrated training: zero human labels, custom seg models at 30+ FPS on edge
iamrobotbear · x · 2026-09-08
What happened
LearnOpenCV used GPT-6 Astra Ultra as an orchestrator to train two segmentation models (RF-DETR-Seg-M) with zero human-supplied labels, from one lazy prompt. Held-out tests show Astra-generated masks approach SAM3 quality.
The workflow
- Add RoboFlow MCP via claude mcp add -s user roboflow
- Ask the LLM to build a dataset, label video/images, and train an RF-DETR model
- Deploy: the new model runs 30+ FPS on a cloud endpoint or on consumer/edge hardware
The thesis
Use a large, slow (1T+ parameter) model that knows a lot to train a purpose-built, fast, small model for edge deployment — framed as the future of realtime vision models.
More from coding & agent
- Gergely Orosz: Local agents launching apps on my desktop need to move to VMs or the cloud — rseroter · 2026-09-08
- An agent skill pushes Blender-generated 3D worlds straight to the Hugging Face Hub — mervenoyann · 2026-09-08
- Kimi K3 detects concurrent file edits, pauses to ask user — who was the human himself — serious_mehta · 2026-09-08
- Ethan Mollick let GPT-6 Astra autonomously build his second brain over 5 days — tomcrawshaw01 · 2026-09-08
- OpenAI revamps Codex use cases page with suggested workflows — umesh_ai · 2026-09-08
- Same prompt, different philosophies: Fable 5.1 ships 992 sprite frames vs GPT-6 Astra's 16-pose sheet — bobo-the-merciful · 2026-09-08