Gemini 3 Pro: the frontier of vision AI

Original Article ↗ Hacker News Discussion ↗

Launch, Links & Product Positioning

  • Several commenters note broken or internal-only links in the blog post and confusing “see in AI Studio” prompts.
  • Some confusion over branding: Gemini 3 Pro (reasoning + vision) vs Nano Banana (image generation) vs other variants; users find the alphabet soup hurts trust and expectations.
  • A few point out that, functionally, this is more a showcase of Gemini 3’s vision abilities than a truly new model.

Benchmarks & Vision Capabilities

  • ScreenSpot-Pro benchmark scores impress many: Gemini 3 Pro ~73% vs Claude Opus 4.5 ~50%, Gemini 2.5 ~11%, GPT‑5.1 ~3.5%, suggesting a large leap in GUI grounding and screen understanding.
  • GPT‑5.x is widely reported as weak at OCR and high-res UI tasks, likely due to aggressive downscaling and token limits; earlier GPT‑4 was seen as better.
  • Commenters see a “data flywheel”: better OCR → more usable scanned books/documents → better models.

Real-World Experiments & Use Cases

  • Users report strong performance on:
    • Complex OCR (including puzzles and timestamp-based letter extraction) where other models failed.
    • Electrical drafting workflows (reading PDFs, mapping outlets into Revit, using code tools).
    • Plant health assessment via live camera.
    • Detailed video descriptions (e.g., Zelda and Witcher gameplay) and potential for audio-described YouTube.
  • Others compare against Amazon Textract: Textract still wins on handwritten character accuracy, while Gemini wins on context and flexible reasoning.

Image Generation vs Understanding

  • Multiple tests show a gap between “understanding” and “generation”:
    • Prompts like “wine glass full to the brim” often yield ~2/3-full glasses.
    • Nano Banana can sometimes draw 5‑legged dogs or odd objects but fails to recognize them as such later.
    • Word-search highlighting and maze-solving remain brittle: models can solve via code, but one-shot visual editing is unreliable.

Limits: Counting, Novel Configurations & “Cognition”

  • Extensive discussion around failures on:
    • Counting legs on 5‑legged animals, fingers on hands, or designing 13‑hour clocks.
    • Identifying hippocampus in MRI slices or solving mazes directly in images.
  • Some view these as evidence that models are pattern-matchers lacking robust conceptual grounding; others argue this is an efficiency trade‑off and similar to human perceptual biases.
  • Long subthreads debate “hallucination” vs generalization, and whether it’s meaningful to call LLM behavior “cognition” or “intelligence.”

Cloud Dependence, Privacy & Market Dynamics

  • Strong concerns about mandatory cloud use, data harvesting, and reliance on US companies, especially for sensitive corporate or governmental data.
  • Some argue this “centralized AI” market ignores a substantial offline/industrial segment needing on‑device or on‑prem models.
  • Others note that most mainstream users do not care and expect Google’s free, data‑subsidized offerings to be highly competitive.

Jobs, Automation & Broader Impact

  • Vision+tooling is perceived as a key bottleneck for full software and CAD automation; several see Gemini 3 as a big step toward agentic “software genies.”
  • Debate over whether this threatens engineering and drafting roles vs mainly automating repetitive tasks with humans still steering.
  • Turing test, “moving goalposts,” and the gap between marketing claims (“true visual and spatial reasoning”) and edge-case behavior are recurring themes.