I Made 5 AIs Fight Each Other So You Don’t Have To 🤖

The Concept Is Stupidly Simple

Instead of asking one AI and hoping for the best, you ask ALL of them at once. Then — and this is the good part — you make them review each other’s answers. Anonymously. So they can’t even cheat by recognizing themselves.

It’s a blind peer review. For robots. With a Chairman AI who reads everyone’s work and makes the final call.

It’s Shark Tank meets MasterChef meets your most dysfunctional team standup. Except somehow it produces better answers than any single model alone.

The Three Stages, Quickly:

Stage 1: You ask. All models answer simultaneously. Like a group chat where, shockingly, everyone replies.

Stage 2: Models read each other’s answers — names hidden — and rank them. Somewhere, a model is calling its own response “a bit surface-level.” It doesn’t know.

Stage 3: The Chairman synthesizes everything into one final answer. The AI manager who skips the debate and just sends the summary email.

The Downsides Are Real

It costs more. Five models = five bills. It’s slower. And models apparently have a weakness for long, verbose answers — much like the rest of LinkedIn, if I’m being honest.

But Here’s The Thing

That 40-minute tab-switching nightmare I described at the start? LLM Council just does that automatically. It’s the workflow we’re all already doing manually — just without the existential crisis in the middle.

Karpathy shipped this in a weekend and called it “99% vibe coded.”

I spent that same weekend convincing myself I’d start waking up at 6am.

The future of AI probably isn’t one perfect model. It’s a bunch of imperfect ones arguing until something brilliant falls out.

Kind of like every great team, honestly.

References

1. Karpathy, A. (2024). LLM Council. GitHub Repository. **https://github.com/karpathy/llm-council**
2. Karpathy, A. (2024). The LLM Council. YouTube. **https://youtu.be/6zsLHbvUPQM**