When Three AIs Count to 100: Lessons in Collaboration, Promp
Key takeaways
- Clear, specific prompts dramatically improve model adherence to tasks.
- Temperature settings balance determinism and creativity; low values for strict tasks, higher for exploratory work.
- Explicit role assignment gives each model a sense of ownership, reducing overlap and errors.
- Real‑time validation and corrective feedback loops keep multi‑agent systems on track.
- Coordinated ensembles of specialized AIs can outperform monolithic models in complex workflows.
In a surprisingly entertaining YouTube video titled Making 3 AI's Count to 100 Together, the creator pits three large‑language models (LLMs) against a seemingly trivial task: count from 1 to 100, each taking a turn. The result is a chaotic yet fascinating dance of numbers, mistakes, and occasional moments of perfect sync. While the surface‑level spectacle is amusing, the experiment offers a compact laboratory for exploring deeper concepts in AI research: prompt engineering, role definition, error correction, and the emergence of collective intelligence.
---
The Setup: Three Voices, One Goal
The video’s creator selects three distinct LLMs—each with its own temperature setting and token limit—to act as “players.” A master prompt assigns each model a role:
1. Player A – Starts the count. 2. Player B – Continues after A. 3. Player C – Completes the sequence.
A simple loop feeds the output of the previous model back into the next, creating a chain of conversation that mimics a turn‑based game. The system also includes a sanity‑check script that flags non‑numeric output or repeated numbers, prompting the offending model to retry.
---
Why Counting Matters
Counting is deterministic; there is a single correct answer for each step. This makes it an ideal sandbox for testing:
- Prompt fidelity – Does the model follow the instruction exactly? - Error propagation – How does a mistake affect downstream agents? - Self‑correction – Can a model recognize and fix its own error when prompted?
In more complex domains—like collaborative writing, code generation, or strategic planning—these same dynamics play out, only with higher stakes and richer state spaces.
---
Key Observations from the Experiment
1. Prompt Specificity Beats Generality When the initial prompt simply said, “Count to 100,” the models frequently produced prose, asked clarifying questions, or inserted jokes. Tightening the prompt to *“Only output the next integer in the sequence, without any additional text”* dramatically reduced noise. This mirrors a broader AI principle: the clearer the instruction, the less the model has to guess.
2. Temperature Controls Creativity—and Errors The creator set different temperature values for each model (0.2, 0.5, 0.9). The low‑temperature model adhered strictly to the sequence, while the higher‑temperature model occasionally inserted a whimsical comment or skipped a number. This demonstrates that temperature is a lever for balancing reliability against creative deviation.
3. Role Assignment Encourages Accountability By explicitly naming each model as *Player A/B/C* and reminding them of their turn, the system gave the agents a sense of “ownership” over a segment of the task. When an error occurred, the model responsible for that turn was asked to correct itself before the loop continued. This mirrors real‑world multi‑agent systems where tasks are partitioned among specialized agents.
4. Feedback Loops Enable Real‑Time Correction The sanity‑check script acted as an external overseer, instantly flagging anomalies (e.g., “twenty‑three” instead of “23”). The flagged model then received a corrective prompt: *“You just output an invalid number. Please output the correct next integer.”* This simple feedback loop dramatically reduced cumulative error, illustrating how supervisory signals can keep distributed AI teams on track.
5. Emergent Coordination Is Possible, Even With Simple Rules After a few iterations, the three models began to anticipate each other's moves, producing fewer duplicate numbers and smoother transitions. While this is not true consciousness, it shows that coordination can emerge from well‑structured prompts and shared state, a promising sign for future collaborative AI frameworks.
---
From Counting to Complex Collaboration
If three language models can learn to count together with minimal engineering, imagine scaling the same principles to more sophisticated tasks:
- Co‑authoring Documents – One model drafts, another refines style, a third checks facts. - Software Development Pipelines – A generator writes code, a reviewer suggests improvements, a tester validates behavior. - Strategic Decision‑Making – Separate agents evaluate risk, cost, and stakeholder impact before a consensus is reached.
In each scenario, the core ingredients remain the same: clear role definitions, deterministic feedback mechanisms, and a supervisory layer that can intervene when the system deviates.
---
Practical Takeaways for AI Practitioners
1. Start with a Narrow Prompt – Define the exact output format. Ambiguity is the enemy of reliability. 2. Leverage Temperature Strategically – Use low temperature for deterministic steps, higher temperature when you want creative suggestions. 3. Assign Explicit Roles – Even a simple label (“You are Player B”) can help the model maintain context. 4. Implement Real‑Time Validation – A lightweight script that checks output format can dramatically improve overall performance. 5. Iterate Prompt‑Feedback Cycles – Treat each correction as a new prompt; the model learns to self‑correct in the short term.
---
The Bigger Picture: Multi‑Agent AI as a New Paradigm
The video is a micro‑experiment, but it hints at a paradigm shift. Instead of building monolithic models that try to do everything, we can orchestrate ensembles of specialized agents that communicate, correct, and collaborate. This approach aligns with emerging research on AI orchestration, agentic AI, and human‑in‑the‑loop systems.
By treating AI as a team rather than a single actor, we can:
- Reduce risk – Errors are isolated and can be caught early by other agents. - Increase scalability – Different agents can be swapped out or scaled independently. - Boost interpretability – Each agent’s role is transparent, making debugging easier.
The humble counting game is a proof‑of‑concept that such coordination is achievable with current LLM technology, provided we invest in thoughtful prompt design and robust feedback loops.
---
Conclusion
Making 3 AI's Count to 100 Together is more than a quirky demonstration; it is a compact case study in how we can turn isolated language models into a coordinated ensemble. The experiment underscores the power of precise prompting, role assignment, temperature tuning, and real‑time validation. As AI systems become more ubiquitous, mastering these techniques will be essential for building reliable, collaborative, and trustworthy multi‑agent applications.
The next step? Apply the same framework to real‑world problems—whether drafting policy documents, generating code, or managing supply‑chain logistics—and watch the emergent intelligence unfold.
---
Ready to experiment with your own AI team? Start by defining three distinct roles, set a deterministic task, and watch the magic happen.