Andrej Karpathy just built an AI that makes GPT, Claude, Gemini and Grok argue with each other before giving you an answer.

@ai.with.andrew

45.8K followers · usually 1.4K views a post

Views

345.9K

Likes

16.5K

Comments

317

Shares

4.7K

Selling · Other Product

LLM Council

https://github.com/karpathy/llm-council ↗

Why it worked

The video leverages the authority of a well-known figure (Andrej Karpathy) and presents a high-value, novel technical solution to a common AI problem (hallucination) in a fast-paced, visually engaging format.

Summary

The video introduces the LLM Council, an AI system that uses four different frontier models to debate and fact-check each other before providing a final answer. It highlights the efficiency of the system and directs viewers to a GitHub repository to access the code.

Structure

1Introduce the LLM Council concept
2Explain how the multi-model debate system works
3Highlight the benefit of reduced hallucinations
4Mention the rapid development time
5Provide a call to action for the GitHub repo

Product placement

The LLM Council GitHub repository is the central subject of the video.

Call to action

Use this Github repo which is linked down in the pinned comment below.

Original caption

Karpathy just made GPT, Claude, Gemini & Grok ARGUE before answering you 😳 Meet the LLM Council — 4 AI models debate, critique, and fact-check each other, then a chairman model gives one final answer. Andrej Weib coded the whole thing in under 48 hours 🤯 GitHub link in the pinned comment! Follow 100x for more crazy AI stuff 🔥 #AI #GPT #Claude #Gemini #Grok

More from @ai.with.andrew

Their other posts in the index, biggest breakout first.

On-screen text

Andrej Karpathy just built an AI that makes GPT Claude Gemini and Grok argue with each other before giving you an answer this is the LLM Council instead of asking one AI model a question and hoping it does not hallucinate this system sends your question to four different frontier models at once but these models then anonymously critique each other's answers pointing out flaws and ranking responses only after this debate does a chairman model synthesize everything into one high confidence answer think of it like having four of the world's smartest AI systems in a boardroom fact checking to you Andrej Weib coded this full thing in under 48 hours 48 hours which reminds us human focus should be on the vision while AI writes the actual code you can just use this Github repo https://github.com/karpathy/llm-council which is linked down in the pinned comment below and for more such AI content follow 100x

Transcript

Andrej Karpathy just built an AI that makes GPT, Claude, Gemini and Grok argue with each other before giving you an answer. This is the LLM Council. Instead of asking one AI model a question and hoping it does not hallucinate, this system sends your question to four different frontier models at once. But these models then anonymously critique each other's answers, pointing out flaws and ranking responses. Only after this debate does a chairman model synthesize everything into one high confidence answer. Think of it like having four of the world's smartest AI systems in a boardroom fact checking to you. Andrej Weib coded this full thing in under 48 hours, which reminds us human focus should be on the vision while AI writes the actual code. You can just use this Github repo which is linked down in the pinned comment below. And for more such AI content follow 100x.