Transcription
Okay, this is big. Researchers at MIT, Stanford, and Nvidia just built an AI agent that communicates telepathically. There's no words or language, just raw thoughts passed between agents.
Now, here's why this matters. Right now, when multiple AI models work as a team, they talk to each other in plain text. For example, agent one thinks, writes its answers out, and then agent 2 reads it, thinks, and then writes back, and so on and so forth. It's basically a group chat between AI agents.
And that sounds fine until you realize that converting thoughts into words and back is slow, expensive, and loses meaning over time. So these researchers had a brilliant idea. Let's just cut out the words entirely. And they did. They called it recursive multi-agent system.
They built a tiny connector called recursive link that plugs directly into each agent's output layers, grabbing thoughts before they even become words and passing them straight on to the next agent. And because there's no words, nothing is lost. You can think of it like a relay race. Each agent builds on the last one's thoughts. Round after round after round, this loops until the final agent delivers the answer. And they didn't even retrain the AI model. Just a tiny connector. That's all.
Now, the results. On competitive level math Olympiad problems, recursive multi-agent systems scored 8% better on average. And on the hardest problems, the gap was even wider. It ran 2.44 times faster and used 75% fewer tokens. Oh, and it cost $4.27 to train. Compare that to roughly $10 for standard finetuning. That is less than half the price while beating it on every benchmark.
So, smarter, faster, and cheaper all three at once. This almost never happens. This one is worth paying attention to. Follow and I'll keep you posted.