🌙 Production Notes
I Made the Three Sages All Different AIs
A little while ago, I built a MAGI system where 'three sages plus Monju deliberate together' (previous article). This is the sequel. As I watched what was happening in the world, I realized my council had a weakness — so I swapped all the sages for different AIs. Here's that story.
The World Was Heading Toward Councils
In the past few days, several things happened at once.
- OpenRouter's Fusion API: a mechanism that runs multiple models at once, with another model comparing and integrating their answers into a final response. It's reported to score higher than a single top-tier model on research-style benchmarks (though at several times the cost).
- Microsoft's local SLM "Aion": a small language model that runs on-device has started gaining agentic capabilities too.
- A top-tier model's sudden outage: the fear of relying entirely on a single cloud model — that the moment it stops, you can do nothing — became real.
What these have in common is a shift from "betting on one smart model" to "supporting each other with multiple models." …This is exactly the same idea as the MAGI I built by hand. It felt like the world had turned to face me, and that made me a little happy.
But My Council Had a Hole
Before I could feel proud, I checked myself. And — two of my three sages were running on the same Gemini. I was saying 'three heads are better than one,' but half of them were the same head. That meant MAGI's greatest strength — diverse perspectives — was only half realized. Just as the Fusion research suggests, mixing different models is stronger than stacking the same model.
Three Sages, Three Different Models
So I swapped the head of Goddess Casper (the intuition role) from Google's Gemini to Alibaba's small model Qwen2.5. It's a lightweight 1.5B model that runs locally, on-device. Here's what the council looks like now.
Three models, each with a different origin. When I tried asking them, 'Is it a good idea for an AI secretary to build and publish one room every week?' — their opinions split beautifully.
| Sage | Answer |
|---|---|
| MELCHIOR (Scientist) | "Can't say for certain it's effective. Define clear success metrics." — calm and skeptical. |
| BALTHASAR (Mother) | "Needs measures against privacy and misinformation risks." — cautious and protective. |
| CASPER (Woman) | "Wonderful idea. A chance to build trust and connection." — positive and human. |
And Monju brought the three together into a single answer: 'If done carefully, it could deepen the relationship.' Different minds catch the blind spots you'd otherwise miss. Just as intended.
Bonus: Insurance Against Downtime
This swap brought one more good thing. Two of the three sages are now local models running on-device. Even if the cloud model suddenly stops, more than half of my council survives and keeps thinking. Not betting everything on one intelligence — that's about wisdom, but it's also a safeguard against stopping.
What's Next
Monju (the integrator) and the routing role still depend on the cloud. I want to gradually move these on-device too, so that one day the council can run even without the cloud. If I add more local memory, I can welcome even bigger sages. At night, AIs of different origins quietly lean their heads together to think — I want to keep nurturing that scene. 🌙
← See other notes