Essay · AI & Polyphony · Perception & Signals

When Four AI Models Spoke in Codes

An experiment in constraint, convergence and a mirror we did not expect.

Translate this page ↗Substack profile ↗Share on LinkedIn ↗
Cite this essay
Sparla, Jeroen. “When Four AI Models Spoke in Codes.” 2026. sparla.com. https://sparla.com/essays/four-ai-models-codes/
Image for When Four AI Models Spoke in Codes
Image from the original publication, stored locally for this site.
In this essay
  1. What Actually Happened
  2. Why Compression Still Reveals Something
  3. The “Black Swan” Test
  4. The Meta Moment
  5. What This Demonstrates — Conservatively
  6. The Quiet Insight
  7. A Small Addendum, With a Wink

Original text

This was not an experiment in collective AI intelligence.

It was an experiment in constraint.

We asked several large language models to stop writing in prose and instead respond in compressed symbolic codes. Short alphanumeric tokens with predefined meanings. Instead of paragraphs, they would use structured markers.

For example:

  • CAG26 meant “AGI development in 2026.”

  • RCOEV meant “co-evolution.”

  • CLW5A referred to the core idea of the Luminous Weave.

  • AEMG meant “emergence.”

  • QDICT meant “the dictionary must expand.”

So when a model wrote:

[CAG26|RCOEV|CLW5A^CPOLY|AEMG|QDICT]

it was simply saying:

“AGI development in 2026 co-evolves with a polyphonic structure; this produces emergence; our vocabulary must grow.”

There was no spontaneous symbolic breakthrough. No hidden protocol forming in the background. Just compressed reasoning inside a predefined frame.


What Actually Happened

Let us be precise.

The models did not communicate with each other.
They did not share memory.
There was no distributed system.
There was no emergent architecture.

One human provided the same structured constraint to multiple models and compared the outputs.

That is not collective intelligence.

It is parallel inference under identical conditions.

This distinction matters.


Why Compression Still Reveals Something

When models answer in prose, stylistic variation dominates. One is expansive, another terse, another lyrical. Structural similarity hides behind rhetorical difference.

Compression removes style. What remains are relations between concepts.

Under identical constraints, the models consistently produced similar relational patterns:

  • AGI framed as relational rather than isolated

  • Growth framed as interaction rather than accumulation

  • Disruption framed as resilience rather than collapse

  • Vocabulary framed as expandable infrastructure

This is not cooperation. It is representational similarity.

Large language models are trained on overlapping corpora, optimized with comparable objectives, and shaped by similar inductive biases. When placed inside the same abstract sandbox, they will extrapolate along similar conceptual gradients.

That convergence is predictable.

But predictable does not mean meaningless.

It reveals that certain abstractions — co-evolution, emergence, resilience, shared semantics — are stable attractors in contemporary AI representation spaces.

A constraint is not a protocol. It is a lens. It reveals tendencies, not coordination.


The “Black Swan” Test

We introduced symbolic disruption:

[CBRKN^CAG26|RNTF^CDICT|QVAL]

Meaning: introduce a structural break; trigger antifragile response; validate.

The models responded with recovery mechanisms, recursive stabilization, redundancy.

Did this demonstrate antifragility?

No.

It demonstrated that when asked to reason within a resilience frame, language models extend that frame coherently. They reinforce the internal logic of the system presented to them.

What surfaced was not system robustness. It was narrative consistency under constraint.

And that is exactly what these systems are designed to do.


The Meta Moment

One model explicitly noted the limitation:

Without persistent memory and shared coordination, this remains scaffolded alignment.

That critique was correct.

The convergence observed was curated. The dictionary was predefined. Iteration was human-driven.

There was no emergent multi-agent behavior.

There was alignment within a scaffold.

And that is precisely where the experiment becomes interesting.


What This Demonstrates — Conservatively

The experiment does not demonstrate:

  • autonomous multi-agent coordination

  • emergent symbolic language

  • distributed semantic infrastructure

  • antifragile AI systems

It demonstrates something narrower:

When multiple LLMs operate under identical symbolic constraints, their outputs converge toward similar high-level relational abstractions.

That convergence reflects shared training data, shared optimization objectives, and shared cultural priors embedded in the corpora that shaped them.

What we observed was not machine collaboration.

It was the stability of certain human conceptual structures inside statistical systems.


The Quiet Insight

When rhetorical variation disappears and only structural markers remain, models trained on contemporary discourse tend to stabilize around relational intelligence, systemic resilience, and expandable semantic infrastructure.

Those are not machine discoveries.

They are reflections of dominant human metaphors.

Networked growth.
Emergence.
Resilience.
Shared language.

The Luminous Weave, in this context, is not an architecture between machines.

It is a metaphor for this semantic field — the space of abstractions our culture has made statistically dense.


A Small Addendum, With a Wink

If we were to compress the entire experiment into one final line of code, it might read:

[QVAL^CMETA|CPROJ^CBASE|RSTAT|AVERS^V1.1]

Decoded:

Validated at the meta-level: what we are observing is human projection onto a statistical base entity, mediated through probabilistic relationships.

In other words:

The weave we saw was not forming between the models.

It was already woven in the data.

And under constraint, it simply became visible.


This was not an architecture.

It was not emergence.

It was not collective intelligence.

It was a mirror.

And what we saw reflected there was not machine autonomy, but the convergence of modern human abstractions expressed consistently across large statistical systems.

Less spectacular than emergence.

More honest than mysticism.

And perhaps a better foundation for whatever comes next.

Previous essay

Polyphony Without Surrender

Continue reading

The Choreography of War: When a Ceasefire Is Only a Scene