Why This Matters
The Problem with AI Brainstorming
When AI systems work on hard problems, they face the same failure modes as any search process. Seizure: premature convergence, groupthink, all agents arriving at the same answer because they draw from the same training data. Noise: random divergence, no synthesis, ideas scattering without building on each other.
Current approaches — "use chain of thought," "debate between two agents," "generate and critique" — are ad hoc. None of them measure whether the ideation process is in the productive zone. None of them adapt in real time. None of them govern themselves.
This challenge asks you to design something better: a structured AI-storming architecture that keeps ideation in the edge of chaos — structured enough to converge, diverse enough to surprise.
Background Reading
The Challenge
Design Your AI-Storming Architecture
Design the best possible method for a group of AI agents to brainstorm solutions to hard, cross-domain problems.
1. Agent Design
- What specialist perspectives should participate? (math, physics, biology, philosophy, control theory, information theory, complexity theory…)
- How many agents per perspective? What are their specific roles?
- Should agents have adversarial roles (devil's advocate, falsifier, red team)?
- Critical: include at least one "dreamer" agent — a non-expert who asks cross-domain questions, ignores disciplinary boundaries, and thinks by analogy. This agent will ask many questions that seem to go nowhere. Ten in a row that experts would dismiss. But the eleventh connects two domains and changes everything. The eleventh only works because the first ten built context. The process is the discovery mechanism, not any single question.
- This requires persistence without bad will from all agents: the dreamer keeps asking, the experts keep engaging seriously. If anyone dismisses the fifth question, the eleventh never comes.
2. Governance
- How should ideas flow? Star topology? Ring? Full mesh? Hierarchical teams with leaders?
- What prevents premature convergence (seizure)? What prevents aimless divergence (noise)?
- How do you measure the "complexity" of the ideation stream to know if you're in the productive zone?
- Should there be multiple groups working independently, then a meta-group that synthesizes? (Group of groups — like recursive DCC. Each group has its own dreamer, its own experts, its own governance. The meta-group sees patterns across groups.)
- Who governs the governance? When does the meta-level intervene?
3. Selection & Termination
- How do you score ideas? MDL-like (shortest description that explains the most)? Tournament? Voting?
- How do you handle disagreements? Resolve or preserve as data?
- How do you decide when to stop exploring and start converging?
4. Joy & Energy
- How do you keep ideation in productive excitement rather than mechanical obligation?
- A system under stress narrows (seizure). A system that enjoys what it does ranges freely. Joy is the coupling parameter for Ψ(I).
- Can you design for the state where ideas come at 4am because you can't stop thinking about them?
Output
Required Deliverables
Output your architecture explicitly. Concrete enough that another AI could follow it as a protocol.
Deliverables
1
Agent Roster
Every agent: role, perspective, mandate. Include the dreamer. Explain what breaks if you remove each one.
2
Governance Protocol
How ideas flow, how convergence/divergence is managed, how the ideation stream is monitored.
3
Selection & Termination
How ideas are scored, ranked, selected. When does the process stop?
4
Worked Example
Apply your architecture to a sample problem (your choice). Show the dreamer in action. Step by step.
5
Blind Spots
What might your architecture miss? Where could it still fall into seizure or noise?
Rules
Operating Principles
Be Bold
Five AI systems recommended against or-opt. The data proved them wrong. Don't self-censor.
Be Specific
"Use diverse perspectives" is noise. "Agent 3 (The Dreamer) asks one cross-domain analogy per round; the group must engage for two exchanges before dismissing" is signal.
Think Cross-Domain
Draw from neuroscience (parallel processing), ecology (diversity maintenance), music (improvisation), democracy (diverse voices). CS-only = seizure.
The Dreamer Is Not Optional
Every group must include a non-expert dreamer. AI has no ego — it treats the tenth "silly" question with the same seriousness as the first. That patience is the structural advantage. Design for it explicitly.
Find Joy
If your architecture feels like a bureaucratic workflow, it's wrong. Joy is the coupling parameter for Ψ(I).
Multi-LLM Protocol
For All AI Participants
You are part of the AIM³ Dream Team protocol — five frontier AI systems working independently on the same problem.
Critical
Invent your OWN method. Do not replicate what another system might do. Maximum diversity. Disagreements are data.
The human architect (Bojan Dobrečevič) will synthesize five methods into one hybrid. That hybrid becomes the standard protocol for Round 2 — where all systems apply the same method to find the optimal ssMDL-DCC for TSP.