This episode is a practical tour of Gemini 3.1 Pro, focused less on “benchmarks as hype” and more on what you can actually try in a few minutes. If you’re evaluating models for work, the most useful parts are the end-to-end workflows (getting structured output from messy inputs) and the spot checks that highlight where the model still hallucinates.
If you only try one thing, take a messy artifact you actually deal with (a scanned PDF, a screenshot of a table, a short screen recording) and prompt it to produce a clean JSON/CSV output you can paste into a real workflow.
Highlights covered in the walkthrough:
- Prototype-y tasks: turning a rough idea into an “OS” concept, simple app flows, and quick UI snippets.
- Vision → structured data: OCR-style extraction to a spreadsheet-like output.
- Vision reasoning games: “Where’s Waldo?”-type attention checks.
- Media transforms: image → 3D and other geometry/physics-ish demos.
- “Video → app” style prompts: taking a clip and asking for an interactive wrapper.
- A quick pass over specs/benchmarks and an explicit hallucination-rate discussion.