When removing spaces broke Nora’s brain

Understanding the mind-boggling inner workings of LLMs shows exactly why AI can be simultaneously smart and stupid at the same time

As always, we’ve been tinkering with our product recently. We created a Neople named Nora as our testing grounds.

The simple task that should have been… simple

Last week, we asked Nora to do something incredibly basic: take a phone number like “123 456 789” and turn it into “123456789”. Just remove the spaces. That’s it.

Any programmer could write this in one line. Heck, any human could do it in seconds. Even the simplest computer program from the 1970s could handle this without breaking a sweat.

Nora? She struggled. Hard.

And before you think “well, your AI must be broken” – nope. This is actually a fascinating glimpse into how modern AI “thinks” differently than both humans and traditional computers. Let me explain why removing spaces can be harder than writing poetry for an LLM.

The weird keyboard in Nora’s brain

To understand what happened, let me show you how Nora actually “sees” text. It’s… weird.

When you see “Hello world”, you see 11 characters: H-e-l-l-o-space-w-o-r-l-d.

When Nora sees “Hello world”, she might see something like: ⊕ ∇

That’s it. Two symbols. The ⊕ means “Hello” and ∇ means “ world” (with the space attached!). She can’t see inside these symbols – she just knows that ⊕ represents that specific sequence of characters.

Think of it like this: Nora has a magical keyboard where each key shows a strange symbol instead of letters. She’s memorized that ⊕ means “Hello”, but she can’t see the H, the e, the l’s, or the o. She just knows: press ⊕, get “Hello”.

The chaos of Nora’s keyboard

Before we get to our space-removal disaster, let me show you just how weird Nora’s keyboard really is. Take the word “apple”:

• “apple” = ℧

• “ apple” (with a space) = ⟚

• “Apple” = ☊

• “APPLE” = ✠

• “apples” = ℸ

• “apple’s” = ⊡

Six variations of “apple” – six completely different buttons in completely different places on her keyboard! To you, all different spellings of “apple” are clearly related. To Nora, ℧ and ☊ are as different as “car” and “philosophy.”

Let’s try Nora’s challenge

Here’s what Nora saw when we asked her to remove spaces from “123 456 789”:

What you see: “123 456 789” What Nora sees: ◊ ⟐ ✦

Where:

• ◊ = “123”

• ⟐ = “ 456” (space included!)

• ✦ = “ 789” (space included!)

Now, we asked her to produce “123456789”. But on her keyboard, that might be a completely different symbol, like ※. Or worse – it might be multiple individual keys.

Can you remove the spaces from ◊ ⟐ ✦? You can’t look inside ⟐ to find the space. You can’t split it apart. You just know ⟐ means “ 456” as one unbreakable unit.

Why this breaks everything

This is why removing spaces is so hard for Nora. She literally cannot see the individual characters inside her tokens. When we say “remove the spaces,” it’s like showing someone this:

“Please remove all the red color from ▣”

You know ▣ contains red (I just told you!), but you can’t see it. You can’t reach in and extract it. It’s just… ▣.

For a traditional computer, text is like a string of beads:

• 1-2-3-[space]-4-5-6-[space]-7-8-9

• Just remove the [space] beads!

For Nora, text is like sealed capsules:

• [capsule1] [capsule2] [capsule3]

• The spaces are locked inside capsules 2 and 3

• She can’t open the capsules

The poetry paradox

Here’s the beautiful irony: Nora can write sonnets, create complex email campaigns, and even debug code. But removing spaces? That’s where she stumbles.

Why? Because she’s incredible at knowing which symbols go together to create meaning. Through millions of examples, she’s learned which symbols together create beautiful prose. She doesn’t need to see inside the symbols – she just needs to know the patterns.

But asking her to manipulate the characters inside those symbols? That’s like asking a pianist to play a song while wearing oven mitts. They know the music, they know the keys, but they can’t feel the individual notes.

What this means for vibe working

This space-removal incident revealed something profound: Nora experiences text in a fundamentally different way than we do.

And that’s actually perfect. We don’t need another character-by-character processor – we have plenty of those. We need someone who understands meaning, context, and intent. Someone who can see the forest, even if she can’t examine each leaf.

When Nora needs to do precise character manipulation, we give her tools designed for that task. It’s not a weakness – it’s just a different way of understanding language.

The upside-down world of AI

We’re living in an upside-down world where AI can write a heartfelt apology email but struggles to count the letters in “strawberry.” Where it can explain quantum physics but trips over removing spaces.

This isn’t a bug – it’s a fundamental part of how LLMs work. They see language as patterns of symbols, not sequences of characters. They understand meaning, not mechanics.

Imagine if you could only read books where every word was replaced with a unique symbol. You’d memorize that ℘ means “love” and ⟁ means “beautiful,” and you could still understand stories and write poetry. But if someone asked you to count the vowels in ℘? Impossible. You know it means “love,” but you can’t see the l-o-v-e.

That’s Nora’s world. And honestly? It’s exactly the perspective we need for a digital colleague who understands what you mean, not just what you type.

---

Curious about more AI quirks and breakthroughs? Follow our journey at neople.io