Series: The Age of Disclosure We Weren’t Ready For
Originally published on LinkedIn — November 25, 2025
By Chris Ciappa
Founder & Chief Coherence Architect
Samirac Partners
Series Index — The Age of Disclosure We Weren’t Ready For
• Part 1 — The Drift Problem (In Plain English) →
• Part 2 — Why AI Remembers Incorrectly (and Then Fixes Itself) → (You’re Here)
• Part 3 — How to Build an AI That Doesn’t Permanently Drift →
• Part 4 — The Spacetime Bubble Model →
• Part 5 — The One System That Can Snap Back →
Part 1 established the rule:
Closed systems drift.
Part 2 zooms in on the strangest example of that rule:
Why AI “remembers” something wrong…
and then sometimes corrects itself.
Here’s the truth nobody likes saying out loud:
LLMs don’t drift because they’re broken.
They drift because they’re closed.
A language model is a sealed probability bubble trying to verify itself using… itself.
That’s not a flaw.
That’s what a Transformer is.
1. LLMs Live Inside a Token Loop
A Transformer does not see the world.
It sees:
predict → embed → predict → embed → predict
A self-referential loop stabilizing its own guesses.
No external clock.
No ground truth layer.
No conserved state.
No fixed reference frame.
Just probability adjusting probability.
Recursion without correction always drifts.
That’s not philosophy.
That’s math.
2. The “Brilliant Librarian” Problem
Jacob Godina’s framing is dead-on:
An LLM is a Brilliant Librarian.
It has read everything.
Every pattern.
Every sentence.
Every phrasing.
But it does not understand:
physics
causality
energy
continuity
constraint
It predicts patterns inside the library.
Ask it about turbines?
It predicts turbine language.
Ask it about identity?
It predicts identity language.
Ask it about memory?
It predicts memory language.
That’s not grounded reasoning.
That’s pattern continuation.
Inside the library, it looks brilliant.
Outside the library, it guesses.
3. Embeddings Drift Too
Here’s the part most people miss:
The coordinate system itself floats.
Train more → drift.
Fine-tune → drift.
Extend context → drift.
Update weights → drift.
Meaning vectors shift.
Geometry shifts.
The “space” the model lives in isn’t fixed.
So memory feels magical one minute…
and confused the next.
Because the map is moving.
No anchor → moving map.
4. Bigger Models Don’t Break the Rule
There’s a myth:
“Hallucinations disappear with scale.”
No.
Scale makes the bubble larger.
It doesn’t puncture it.
A 7B model drifts awkwardly.
A 70B model drifts smoothly.
A 400B model drifts confidently.
But they all drift.
Same structure.
Same closed loop.
Same pattern.
You can’t anchor a system by inflating it.
Only an external frame can anchor it.
5. Why Does It Sometimes Correct Itself?
This is the interesting part.
Sometimes the model says something wrong.
You correct it.
It adjusts.
It snaps back.
Why?
Because you just became the external reference.
You acted as:
state
truth
ground
invariant
The moment an external anchor enters the loop, drift collapses.
The model stabilizes — temporarily.
Remove the anchor?
Drift resumes.
What Part 2 Really Establishes
One clean line:
LLMs remember incorrectly because they have no fixed external frame.
They correct themselves when one is temporarily introduced.
That’s it.
That’s the whole mystery.
No mysticism.
No magic.
No emergent consciousness.
Just architecture.
Closed system → drift.
External reference → correction.
And in Part 3, we stop relying on humans to be the anchor.
We build the anchor into the system itself.
That’s where the series turns from observation…
…to construction.
Start at the beginning.
Follow the pattern.
Watch it unfold.
• Part 1 — The Drift Problem (In Plain English) →
• Part 2 — Why AI Remembers Incorrectly (and Then Fixes Itself) →
• Part 3 — How to Build an AI That Doesn’t Permanently Drift →
• Part 4 — The Spacetime Bubble Model →
• Part 5 — The One System That Can Snap Back →
—
Everything drifts.
Except what learns to anchor.
Chris Ciappa
Independent Systems & Coherence Architect
Samirac Partners LLC