🍷 Follow Up: Nano Banana 2 - Wine Glass Test

Michael Wacht

21d (edited) • 🟢 AI Enthusiast

This is a follow-up to my original “Wine Glass Test” — a simple experiment that turned into something more interesting.

After my first post, I received a thoughtful suggestion from

@Matthew Sutherland

His advice was straightforward:

Be more prescriptive.

So I refined the prompt to this:

“Create a glass of wine that is full, red wine. It needs to be at the brim, so not to run over, and not below the brim to show any space between the brim and the surface of the wine in the glass.”

The image below is the direct result.

And the result is telling.

🍷 What This Actually Proves

This wasn’t about aesthetics.

It was about bias and instruction.

When I originally asked for a “full glass of wine,” the model produced what most restaurants would call full — but still left space at the top.

That’s not an error.

That’s statistical bias.

The model leaned into the most common interpretation of “full.”

When the instruction became extreme and structured, the behavior changed. It complied precisely.

🍷 There are two observations that I see with this test:

1️⃣ Prompting Is a Skill

We often talk about model bias as if it’s a flaw.

It’s not.

It’s probability doing what probability does.

My first prompt allowed the model to default to “standard pour.”

The refined prompt removed ambiguity.

By defining the boundary conditions — no gap, no overflow — the model had to break from its average tendency and execute exactly.

That’s not luck.

That’s instruction design.

Prompting isn’t just writing a sentence.

It’s mapping expectation into structure.

And as Matthew pointed out, that skill develops iteratively.

2️⃣ Natural Language Still Has Friction

The deeper takeaway isn’t that the model can create a perfectly full glass.

It’s that everyday language is still ambiguous to it.

When a human says “full glass of wine,” we infer intent through context.

The model infers through probability.

Those are not the same.

For AI to feel seamless in daily life, we shouldn’t need to mathematically define “full.”

But today, precision still requires precision.

🍷 The Real Distinction

As a business tool, Nano Banana 2 performs extremely well when given structured, constraint-based instructions.

But as an everyday assistant, it still exposes the gap between statistical understanding and semantic understanding.

That gap is where the next layer of AI progress will happen.

The interesting part of this test isn’t the wine.

It’s the friction between language and logic.

And every time we refine a prompt, we’re not just training the model.

We’re training ourselves.

10 comments

AI Bits and Pieces

skool.com/ai-bits-and-pieces

Build real-world AI fluency to confidently learn & apply Artificial Intelligence while navigating the common quirks and growing pains of people + AI.

Welcome

Community Rules

AIS+ for AI Builders & AI Agencies

Leaderboard (30-day)

+88

+43

+18

+18

+12