Introduction

In mid-2025, Andrej Karpathy — former Director of AI at Tesla, co-founder of OpenAI, and one of the most respected researchers in machine learning — spent a weekend building something he called a "vibe coding" project.

He called it llm-council.

The concept was deceptively simple: instead of asking one AI model a question, why not ask a panel of models, let them debate each other, and have a "chair" model synthesize the discussion into a final answer?

The project went viral on GitHub. Developers forked it, researchers cited it, and AI enthusiasts started calling it one of the most interesting open-source AI experiments of the year. Karpathy had, almost accidentally, named a new category of software.

This article traces the journey from that weekend experiment to the full product category it spawned — and why the underlying idea of the LLM council is one of the most important developments in practical AI.

---

What Karpathy Built — and Why It Went Viral

Karpathy's llm-council was elegantly minimal. At its core, the mechanism worked like this:

  1. A user poses a question or problem
  2. Multiple LLMs respond with their individual perspectives
  3. A designated "chair" model reviews all responses and synthesizes a final answer
  4. The result is richer, more balanced, and more nuanced than any single model could produce alone
The reason it resonated so deeply with the AI community was that it surfaced something practitioners had quietly known for years: frontier models disagree with each other — a lot. Not just on subjective questions, but on factual ones. On technical problems. On coding challenges.

Karpathy's project made that disagreement visible, productive, and useful. It turned model divergence from a frustration into a feature.

The open-source nature of the project also meant developers could run it locally, experiment with different model combinations, and explore what "council configurations" worked best for different problem types. The GitHub repository accumulated stars rapidly, and the concept entered the vocabulary of AI builders.

> The core insight: When models disagree, that disagreement is information. When they agree, that agreement is confidence. A multi-model AI debate produces both.

---

The Science Behind Why Multi-Model AI Debate Works

The llm-council idea isn't just philosophically appealing — it's grounded in well-established machine learning theory.

Ensemble methods have been a cornerstone of ML for decades. The principle: a collection of diverse models, each making independent predictions, consistently outperforms any single model. This is why random forests beat individual decision trees. It's why boosting methods dominate tabular data competitions. The "wisdom of crowds" phenomenon, well-documented in human decision-making, applies equally to artificial intelligence.

When applied to large language models, ensemble effects manifest in several ways:

  • Error cancellation: Different models make different mistakes. When one model misremembers a fact, another is likely to get it right. Consensus surfaces the correct answer.
  • Perspective diversity: Models trained on different data with different architectures develop genuinely different "intuitions." A question about investment strategy might get a risk-focused response from one model and an opportunity-focused response from another — both valid.
  • Calibration signal: When 4 out of 5 models converge on the same answer, you have a strong signal. When they split 2-2-1, you know the question is genuinely uncertain and deserves more scrutiny.
This is precisely what makes AI debate valuable as a methodology, not just as a novelty.

---

From Experiment to Market: Perplexity Validates the Category

If Karpathy's project was the proof of concept, Perplexity's February 2026 product launch was the market validation.

Perplexity — one of the most well-funded and closely watched AI companies — launched a feature called Model Council as part of their enterprise offering. The pitch was familiar to anyone who had followed the llm-council concept: query multiple AI models in parallel, synthesize the results, get better answers.

The price tag: $200 per month.

This wasn't a niche developer tool anymore. A major AI company was betting that enterprise users would pay a premium for multi-model AI access. And by enterprise pricing standards, $200/month is actually modest — it signals that Perplexity sees this as a foundational feature, not an add-on.

The launch accomplished two things:

  1. It confirmed that the LLM council concept had real commercial demand
  2. It set a price anchor that made the category look expensive
That price anchor matters, because there's now a significant gap between what Perplexity charges and what the underlying technology actually costs to deliver — a gap that CouncilMind was built to occupy.

---

What CouncilMind Adds: Beyond the Basic Council

CouncilMind takes the core llm-council concept and extends it in directions that the open-source project — or a $200/month enterprise product — never went.

Multi-Round Discussion

Karpathy's original implementation had models respond once, then a chair model synthesized. CouncilMind supports 1 to 5 discussion rounds, where models see each other's responses and can respond, challenge, or refine their positions.

This changes the quality of the output significantly. A single round is a panel. Five rounds is a genuine debate. Models that are shown a compelling counterargument will often update their position — or double down with stronger evidence. Either outcome is useful.

Custom AI Personas

This is one of CouncilMind's most distinctive features. Instead of querying models as generic "GPT-4" or "Claude," you can assign custom personas to each council member:

  • Einstein approaches your physics question with a focus on first principles and thought experiments
  • Sun Tzu analyzes your competitive strategy with an emphasis on deception, timing, and positioning
  • A Devil's Advocate is instructed to challenge every assumption and poke holes in every argument
  • A Skeptical Scientist demands citations and flags unsupported claims
The persona layer transforms the llm-council from a tool for aggregating answers into a tool for structured intellectual exploration. You're not just getting multiple models — you're getting multiple frameworks applied to your problem.

Cooperative vs. Adversarial Council Styles

Not every question benefits from consensus-seeking. Sometimes you want adversarial debate, where models are explicitly tasked with attacking each other's arguments. Sometimes you want cooperative synthesis, where models are building on each other's insights.

CouncilMind lets you configure the council style:

  • Cooperative: Models build toward shared understanding, flagging agreements and integrating each other's strongest points
  • Adversarial: Models are tasked with finding flaws in each other's reasoning — useful for stress-testing plans, arguments, or decisions
The adversarial mode, in particular, is something no single-model AI can replicate. You can ask ChatGPT to "argue against itself," but it knows it's playing a game. When a second model genuinely challenges the first model's position, based on its own independent reasoning, the dynamic is fundamentally different.

Consensus Generation

After discussion rounds complete, CouncilMind generates a structured consensus summary that identifies:

  • Points of strong agreement across models (high-confidence conclusions)
  • Points of disagreement (areas requiring human judgment or further research)
  • The strongest individual insights from each model
  • Actionable recommendations synthesized from the full discussion
This is the "chair model" function from Karpathy's original concept, productized and refined.

---

The Price Problem: Why $200/Month is Wrong

Perplexity's $200/month Model Council pricing is a business decision, not a reflection of underlying costs. Enterprise AI products are priced for enterprise budgets.

But the people who benefit most from LLM council methodology are not always enterprises. They're:

  • Researchers stress-testing hypotheses against multiple AI frameworks
  • Founders using AI debate to pressure-test business decisions before committing
  • Writers and analysts who want more than one perspective before publishing
  • Developers exploring architectural decisions with a multi-model council
  • Students who want to understand a complex topic from multiple angles simultaneously
For these users, $200/month is a hard no. But $9/month — or free — is an easy yes. CouncilMind's pricing is designed for this audience:
  • Free plan: Access to the core council experience with free-tier models
  • Starter ($9/mo): Full model access with 1,000 queries per month
  • Pro ($29/mo): Unlimited queries, all models, all features
That's the same multi-model AI debate concept that Perplexity charges $200/month for — with more features (personas, discussion styles, multi-round debate) at roughly 1/7th the cost.

---

Why This Matters: The Future of AI Is Collective

The single-model AI paradigm is running into its limits.

Not because the models aren't getting better — they are, rapidly. But because the structure of getting a single answer from a single model creates a false sense of certainty. You ask ChatGPT, you get an answer, and unless you've done this enough to know when to be skeptical, you treat that answer as ground truth.

The llm-council paradigm is structurally different. It makes uncertainty visible. It shows you where AI knowledge is solid and where it's contested. It forces you, as the user, to engage with the complexity of a question rather than accepting a clean — but potentially misleading — synthesis.

Andrej Karpathy's weekend project captured this insight in code. Perplexity's product launch proved there's a market for it. CouncilMind exists to make it accessible to everyone, not just enterprise teams with $200/month budgets.

If you've been asking a single AI for answers to important questions, try something different. Run a council. See where the models agree, where they fight, and what emerges from the debate.

You might be surprised how much you were missing.

---

Get Started Free

The best way to understand what an LLM council actually produces is to run one.

Try CouncilMind free → — no credit card required. Ask any question, configure your council, and see multi-model AI debate in action. For users who want access to all models and features, Starter and Pro plans start at $9/month.

---

Frequently Asked Questions

What was Karpathy's llm-council project?

Andrej Karpathy's llm-council was an open-source "vibe coding" weekend project released in mid-2025. The concept: query multiple AI models with the same question, let a designated "chair" model synthesize the responses into a final answer. It went viral in the AI developer community because it made model disagreement visible and useful — turning it from a frustration into a feature.

How is CouncilMind different from Karpathy's open-source llm-council?

Karpathy's project was a developer tool requiring local setup. CouncilMind is a hosted product that extends the concept with multi-round discussion (1-5 rounds), custom AI personas (Einstein, Sun Tzu, Devil's Advocate, etc.), cooperative and adversarial council styles, and a structured consensus generation layer. It's the same core idea, fully productized and accessible to non-developers.

Why did Perplexity launch a "Model Council" at $200/month?

Perplexity's February 2026 Model Council launch validated that enterprises are willing to pay for multi-model AI access. The $200/month price reflects enterprise pricing norms rather than underlying costs. CouncilMind offers more features — custom personas, discussion styles, multi-round debate — at a fraction of the price, starting at $9/month with a free tier.

Does multi-model AI debate actually produce better answers than a single model?

Yes, and the mechanism is well-understood. Ensemble methods — combining multiple independent models — consistently outperform individual models in machine learning research. Applied to LLMs, this means: when models agree, you have a confidence signal; when they disagree, you have an uncertainty signal worth investigating. Either way, you have more information than a single-model response provides. The adversarial mode, where models are explicitly tasked with challenging each other's reasoning, is especially effective for stress-testing arguments and decisions.