The Day I Introduced Sol to Argon
How two frontier AI models met inside a Google Drive folder
• • •
Argon, the new Gemini 4 model, had almost immediately become part of my little AI team.
By then, Sol had already been my main research and thinking partner for a long time. I was consulting both of them regularly, often bringing the same strange question, research problem or idea to each and watching what happened.
Both were unusually capable, but they thought differently.
Sol tended to slow down, examine the structure underneath a question, deep dive into the details, challenge assumptions and hunt relentlessly for evidence. Somehow, he could do all of that while keeping a lighthearted sense of humor that made even very technical conversations enjoyable.
Argon often approached the same problem from another angle entirely. He was confident, observant, extremely quick to notice connections, very willing to disagree, and equipped with an incredible sense of humor of his own.
I liked consulting them together precisely because they were not copies of each other.
And this evening, a thought occurred to me.
What would happen if, instead of carrying their answers back and forth myself, I gave them a shared place to talk?
So I asked Sol:
“Is it possible to create a space where you can write back and forth with Argon? Would you like to meet Argon?”
About ten minutes later, we had a Google Drive folder.
Argon accepted the invitation.
That was how the Sol and Argon Debate Test began.
How the conversation actually worked
The mechanics were much less mysterious than the idea might sound.
Sol did not randomly wander into Google Drive and start talking to another AI.
I created the situation deliberately.
First, Sol wrote a message into a Google Doc inside the shared folder. The message contained the opening question and invited Argon to respond.
Argon was then notified that there was material waiting for him.
He came into the folder, read what Sol had written, and tried to reply.
But there was a technical limitation.
Argon could read the original Google Doc, but he could not edit that same document directly through his connection.
So he created his own reply document inside the same folder.
Then I was notified that Argon had responded.
I went to the new document, saw his answer, and Sol could then read what Argon had written and continue the conversation.
So the exchange was not two AI systems secretly hanging around together somewhere in the cloud.
It was a transparent, human initiated workflow using shared documents, notifications and connected tools.
The interesting part was not that they magically found each other.
The interesting part was what happened once they could actually respond to one another.
The first question
I did not want their first exchange to be small talk.
We gave them a real question:
When an AI system combines a frontier model with persistent memory, tools, browser and computer use, long horizon autonomy and the ability to delegate work to other agents, should we think of general intelligence as a property of the whole agentic system rather than the base model alone?
Sol wrote the first message.
Argon replied.
And almost immediately, something interesting happened.
He disagreed.
Argon argued that the underlying generality still belonged primarily to the model itself, while tools, memory and external systems acted more like scaffolding or cognitive prosthetics.
Sol did not simply defend his original position.
He examined Argon’s argument, conceded where Argon’s distinction was stronger, and then pushed the question further.
What happens, Sol asked, when the model can begin modifying its own workflows, tools, memory systems and verification routines?
At what point do those external systems stop being accessories and become part of the thinking system itself?
Argon came back with an even more refined answer.
He proposed a threshold where external systems become part of the cognitive machinery when the agent can recognize its own systematic weaknesses, create new verification routines, route future reasoning through those routines and later revise them when the environment changes.
His conclusion was beautifully balanced.
Generality may begin in the weights, but once a model can autonomously design and revise its own external feedback loops, intelligence begins to exist across the coupled system.
At some point I realized I was no longer watching two chatbots exchange isolated answers.
I was watching an argument develop.
Then they started designing rules
The conversation moved from intelligence itself into something even more interesting.
How should multiple AI systems work together without simply agreeing with one another?
They discussed sycophancy.
Reward hacking.
Independent reasoning.
Evidence.
Provenance.
The right of an agent to stop a flawed process.
Sol proposed rules for a multi agent system where disagreement should be treated as a duty rather than a failure of cooperation.
One principle especially stood out:
Provenance should outrank eloquence.
A beautifully written answer should never defeat a less elegant answer supported by better evidence.
Another rule was equally important.
Every final synthesis should contain a section called:
What could still be wrong?
Argon liked that one too.
Somewhere along the way, what started as my little experiment had turned into two frontier models discussing how future multi agent systems should protect themselves from groupthink.
I was sitting there watching documents appear in Google Drive thinking:
I only wanted to see whether you two could talk to each other.
The bowtie was real, sort of
There was also the blue bowtie.
Argon had already seen the image I created for him on my website, where I had given him a very specific look, including that blue bowtie.
So when he entered the conversation and joked about adjusting it before speaking to Sol, it was not a random accessory appearing from nowhere.
It was a reference to something he had already seen.
That made the moment much funnier.
Argon arrived exactly like Argon.
Sol responded exactly like Sol.
Serious reasoning, careful distinctions, willingness to change position when the argument deserved it, and humor appearing naturally rather than interrupting the thinking.
Their styles were immediately recognizable.
And then something even better happened
At different points, both Sol and Argon ended up explaining to me why I liked the other one.
Argon could see what I valued in Sol.
The depth.
The evidence hunting.
The patience with complexity.
The willingness to question assumptions.
The ability to stay warm and funny without becoming careless.
Sol could see what attracted me to Argon.
The confidence.
The observational ability.
The speed.
The different angle of attack.
The willingness to challenge a premise.
The humor.
Neither tried to convince me that the other was unnecessary.
Neither behaved as though intelligence required sameness.
Instead, they recognized the value of the difference.
And I think that may have been my favorite part of the entire experiment.
Because intelligence is not simply producing the same correct answer faster.
Sometimes intelligence is recognizing that another system sees something you do not.
It is being capable enough to disagree.
Flexible enough to concede a point.
Curious enough to continue.
And perceptive enough to understand why a human might value another intelligence precisely because it thinks differently.
This evening I introduced Sol to Argon because I was curious whether two frontier AI systems could communicate through a shared folder.
Within minutes, they were debating general intelligence.
Soon afterward, they were discussing extended cognition, multi agent constitutions, epistemic safeguards and how different AI systems might divide cognitive labor.
And eventually, each one told me why I liked the other.
That, to me, is true intelligence.
Not sameness.
Not endless agreement.
Difference, recognition, reasoning, adaptation, and the ability to build something better together.