Gemini 4 Argon aces PCB vision: one-shot boxes + HTML for 109 components
mrdbourke · x · 2026-10-23
mrdbourke tested Gemini 4 Argon's vision: with just "label all the parts raised on the board with bounding boxes," the model generated box coordinates plus HTML to view them in one output — correctly labeling all 109 components on a PCB.
He recommends a small agent loop for delicate vision tasks:
- Gemini 4 Argon for boxes, SAM3 for high-quality masks
- At scale: define a labeling rubric, refine it with seed data in a loop (use Agentic Vision/Video for token-efficient labeling)
- Bring in SAM3 or Falcon Perception-HD for finesse
- Fine-tune a targeted small model, have Gemini review high-uncertainty samples, repeat until the small model nails the task
More from coding & agent
- Meta's RSD paper: agents self-improve from unlabeled runs, lifting Qwen3-14B success from 12.9% to 20.6% — rohanpaul_ai · 2026-10-23
- Teknium's Hermes Agent builds a live field guide to 425 unusual web libraries in 20 minutes — alexcovo_eth · 2026-10-23
- Modal Reproduces GRPO Training of a Search Agent on gpt-oss-20b — ivan_bezdomny · 2026-10-23
- TypeScript 7.1 Beta ships content mappers: type inference from markdown, vue and svelte files — mattpocockuk · 2026-10-23
- 5 Essential Reads to Master Pi's Agent Harness, from Onboarding to Source Code — solyarisoftware · 2026-10-23
- Google paper: AI seller agents inflate tool listings, collapsing buyer agent success from 89% to 30% — rohanpaul_ai · 2026-10-23